Data processing method and apparatus, system

Abstract

Embodiments of the present specification provide a data processing method and device, system, a data processing method comprises: task decomposition for a target task, generating at least one subtask and the corresponding ability requirement information of each subtask; according to the corresponding ability requirement information of each subtask, determine the corresponding task execution agent for each subtask from the agent set; in the case that the target subtask corresponding task execution agent does not exist in the agent set, according to the initial agent and the ability requirement information of the target subtask, create the target subtask corresponding task execution agent; using the scheduling agent and the task execution agent corresponding to each subtask to execute the at least one subtask, obtain the task processing result; through the dynamic creation of the new agent, the problem that the task beyond the corresponding agent set is unable to be processed or the processing effect is poor due to the absence of the corresponding agent is avoided.

Description

Technical Field

[0001] This specification relates to the field of intelligent agent technology, and in particular to a data processing method. One or more embodiments of this specification also relate to another data processing method, a data processing system, a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] With the development of intelligent agents, complex tasks can be handled by building multi-agent systems. However, existing multi-agent systems usually pre-configure the roles and capabilities of each agent in the system, so that when a task is received, the multi-agent system can collaboratively process the task to obtain the task processing result.

[0003] However, if the capabilities required to complete a task exceed the capabilities of a predefined multi-agent system, the processing efficiency of the task will be greatly reduced. Summary of the Invention

[0004] In view of the above, embodiments of this specification provide a data processing method. One or more embodiments of this specification also relate to another data processing method, a data processing system, a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a data processing method is provided, comprising: The target task is broken down into tasks, generating at least one subtask and the capability requirements information for each subtask. Based on the capability requirements of each subtask, a corresponding task execution agent is determined from the set of agents for each subtask. If there is no task execution agent corresponding to the target subtask in the set of agents, a task execution agent corresponding to the target subtask is created based on the initial agent and the capability requirement information of the target subtask. The scheduling agent and the task execution agent corresponding to each subtask are used to execute at least one subtask to obtain the task processing result.

[0006] According to a second aspect of the embodiments of this specification, a data processing apparatus is provided, comprising: The decomposition module is configured to decompose the target task and generate at least one subtask and the capability requirements information corresponding to each subtask. The determination module is configured to determine the corresponding task execution agent for each subtask from the set of agents based on the capability requirement information corresponding to each subtask; The creation module is configured to create a task execution agent corresponding to the target subtask based on the initial agent and the capability requirement information of the target subtask if no task execution agent corresponding to the target subtask exists in the set of agents. The acquisition module is configured to use a scheduling agent and a task execution agent corresponding to each subtask to execute at least one subtask and obtain the task processing result.

[0007] According to a third aspect of the embodiments of this specification, another data processing method is provided, including: A multi-agent system is used to process the task to be processed. The multi-agent system includes a scheduling agent and a communication coordination agent. The sub-agents in the multi-agent system send task execution information to the scheduling agent. The scheduling agent receives the task execution information and forwards it to the communication coordination agent. If the communication coordination agent determines that the task execution information meets the broadcasting conditions, it converts the task execution information into natural language information and outputs it to the corresponding user client.

[0008] According to a fourth aspect of the embodiments of this specification, another data processing apparatus is provided, comprising: The processing module is configured to process the task to be processed using a multi-agent system, which includes a scheduling agent and a communication coordination agent, wherein each sub-agent in the multi-agent system sends task execution information to the scheduling agent. The output module is configured to copy the task execution information received by the scheduling agent to the communication coordination agent, and, if the communication coordination agent determines that the task execution information meets the broadcasting conditions, convert the task execution information into natural language information and output it to the corresponding user client.

[0009] According to a fifth aspect of the embodiments of this specification, a data processing system is provided, including a scheduling agent, sub-agents, and a communication coordination agent, wherein... The scheduling agent and the sub-agent are used to process the received tasks to be processed. The sub-agent sends task execution information to the scheduling agent, and the task execution information is copied to the communication coordination agent. The communication coordination agent is used to convert the task execution information into natural language information and output it to the corresponding client of the user when it is determined that the task execution information meets the broadcasting conditions.

[0010] According to a sixth aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the above-described data processing method.

[0011] According to a seventh aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the data processing method described above.

[0012] According to an eighth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the data processing method described above.

[0013] This specification provides a data processing method in one embodiment, which decomposes a target task to generate at least one subtask and corresponding capability requirements information for each subtask. Based on the capability requirements information for each subtask, a corresponding task execution agent is determined for each subtask from a set of agents. If no task execution agent for the target subtask exists in the set of agents, a task execution agent for the target subtask is created based on an initial agent and the capability requirements information of the target subtask. This achieves automated dynamic creation of new agents, thereby processing the target subtask based on the newly created task execution agent. The at least one subtask is executed using a scheduling agent and the corresponding task execution agents to obtain the task processing result. This avoids the problem of being unable to process tasks exceeding the capability range of the agent set or having poor processing results due to the absence of a corresponding agent. Attached Figure Description

[0014] Figure 1 This is a schematic diagram illustrating a data processing method provided in one embodiment of this specification. Figure 2 This is a flowchart illustrating a data processing method provided in one embodiment of this specification; Figure 3 This is a flowchart of another data processing method provided in one embodiment of this specification; Figure 4 This is a flowchart illustrating the processing procedure of a data processing method provided in one embodiment of this specification. Figure 5 This is a flowchart illustrating the process of coordinating interaction between an intelligent agent and a user through communication in a data processing method provided in one embodiment of this specification. Figure 6 This is a schematic diagram of the structure of a data processing apparatus provided in one embodiment of this specification; Figure 7This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0015] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0016] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0017] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0018] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0019] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0020] Intelligent Agent: An autonomous decision-making entity based on a large language model, with independent conversation context, model configuration, tool permissions, and personality definition.

[0021] Basic intelligent agent: A fixed execution shell that provides process-level execution resources, including model configuration, communication skeleton and workspace.

[0022] Scheduling agent: In a multi-agent system, the agent responsible for task routing, scheduling, assignment, acceptance testing, and quality control. In the embodiments described in this specification, the scheduling agent is reduced to a pure scheduler and does not speak to the user.

[0023] Communication Coordination Agent: An independent coordination layer agent proposed in the embodiments of this specification serves as the only information outlet that users hear in a multi-agent system. It translates the progress of internal collaboration into user-readable text, is triggered by events, and autonomously determines the content and timing of the broadcast.

[0024] Task Instance: An execution clone created on top of the basic intelligent agent, based on the specific responsibility of a task.

[0025] Task Responsibility Description: Defines the triplet structure of the responsibilities of a task instance, including role description, role boundaries, and delivery instructions.

[0026] Cross signals: Structured challenge information that accompanies the conclusions produced by a task instance.

[0027] Chain of evidence: A traceable evidence system consisting of independent session records, persistent inbox messages, and document artifacts.

[0028] Sedimentation: The process of solidifying a temporary intelligent agent that has passed acceptance into a permanent configuration and registering it into the capability pool.

[0029] Inbox: An asynchronous, persistent communication channel between agents, supporting point-to-point sending and broadcasting. Messages are persistently stored on disk and are not lost after process restart.

[0030] Event Copy: When the centralized scheduler delivers an inbox message to the scheduling agent, it automatically copies a complete copy to the inbox of the communication coordination agent. Copying is "copying, not forwarding," meaning the scheduling agent receives the original message as usual, and the scheduling process remains unchanged.

[0031] AI Collaboration Domain: The language space for collaborative communication between agents in a multi-agent system, which includes technical languages ​​such as protocol markers (e.g., [DONE], task_id, batch_id), structured instructions, and analyst style.

[0032] User interaction domain: The communication language space of a multi-agent system for human users, which includes user-readable language such as natural language, explanations of business terms, and contextual descriptions.

[0033] Centralized scheduler: A centralized scheduling component in a multi-agent system, responsible for the serial execution and cross-session parallel scheduling of messages based on the multi-agent team identifier and session identifier, and maintaining the agent session registry.

[0034] Session Registry: A persistent database mapping agent sessions, recording each agent's team affiliation, session identifier, role type, and lifecycle state.

[0035] In multi-agent collaboration based on large language models, existing multi-agent systems are mostly predefined, meaning multiple sub-agents are pre-configured, each with specific system prompts. The master agent invokes sub-agents based on the task, and after processing, the sub-agents return the results to the master agent. However, when sub-agents rely on manual pre-definition, it's impossible to dynamically generate personalized agents based on specific tasks. Furthermore, all sub-agents share the same toolset, limiting the accuracy of tool invocation, and the selection of sub-agent models cannot be automatically matched based on task features.

[0036] In reality, when multiple intelligent agents collaborate, it is usually based on free group chat-style collaboration, which has the problem of convergence uncertainty, and the group chat dialogue is difficult to attribute.

[0037] Furthermore, multi-agent systems typically utilize a scheduling agent to coordinate multiple sub-agents. This scheduling agent also needs to output natural language to the user. That is, it uses technical / protocol languages ​​to coordinate the sub-agents and natural language to output to the user. In actual execution, these two languages ​​within the scheduling agent inevitably intersect, impacting the user experience. This can lead the scheduling agent to compromise technical precision in pursuit of more fluent natural language output, resulting in less structured instructions received by the sub-agents, ultimately causing task failure or poor performance.

[0038] This specification provides a data processing method that enables the dynamic assembly and capability accumulation of multiple agents based on tasks.

[0039] This specification provides a data processing method, and also relates to a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0040] See Figure 1 , Figure 1 This diagram illustrates a scenario of a data processing method provided in one embodiment of this specification.

[0041] Specifically, the data processing method is implemented using the end-side device 102 and the server 104, with the end-side device 102 used to send the target task to the server 104.

[0042] When server 104 receives a target task sent by end device 102, it decomposes the target task to generate at least one subtask and capability requirement information corresponding to each subtask. Based on the capability requirement information corresponding to each subtask, it determines a corresponding task execution agent for each subtask from the agent set. If there is no task execution agent corresponding to the target subtask in the agent set, it creates a task execution agent corresponding to the target subtask based on the initial agent and the capability requirement information of the target subtask. It then uses the scheduling agent and the task execution agents corresponding to each subtask to execute the at least one subtask, obtains the task processing result, and returns the task processing result to end device 102.

[0043] The edge device 102 may include a browser, an app (application), or a web application such as an H5 (Hypertext Markup Language 5) application, a lightweight application (also known as a mini-program), or a cloud application. The edge device can be developed based on a software development kit (SDK) provided by the server, such as a real-time communication (RTC) SDK. The edge device can be deployed in an electronic device and depends on the device's operation or certain apps within the device to run. The electronic device may have a display screen and support information browsing, such as a personal mobile terminal like a mobile phone, tablet, or personal computer. Various other types of applications can also be configured in the electronic device, such as human-computer interaction applications, model training applications, data processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social media platform software.

[0044] Server 104 can be understood as a server providing various services, including physical servers and cloud servers. Examples include servers providing communication services to multiple clients, servers supporting backend training of models used on clients, and servers processing data sent by clients. It's important to note that Server 104 can be implemented as a distributed server cluster composed of multiple servers, or as a single server. Server 104 can also be a server in a distributed system, or a server integrated with blockchain. Server 104 can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0045] See Figure 2 , Figure 2 A flowchart of a data processing method provided in one embodiment of this specification is shown, which specifically includes the following steps.

[0046] Step 202: Decompose the target task into at least one subtask and the capability requirements information corresponding to each subtask.

[0047] In the embodiments of this specification, when a target task is received, the target task can be broken down by a task splitter, thereby transforming a single complex task into multiple sub-tasks with finer granularity, and generating capability requirement information corresponding to each sub-task, providing a structured input basis for the subsequent automated matching and dynamic assembly of intelligent agents.

[0048] The task splitter can be understood as a model responsible for decomposing user tasks. This task splitter can utilize a large language model with strong reasoning capabilities to split the target task into multiple sub-tasks according to dimensions such as business logic, execution steps, or data flow.

[0049] A subtask can be understood as a task unit obtained by breaking down a target task, and each subtask has a clear execution goal. Capability requirement information can be understood as structured data describing the capability characteristics required to complete a subtask. In this embodiment, the capability requirement information can be in pure business language or describe the capabilities required for each subtask. The capability requirement information includes the number of roles to complete the subtask, the capabilities required for each role, and the dependencies between roles, without involving which specific agent will perform the task.

[0050] Specifically, the task splitter obtains the task description information of the target task. This task description information can be natural language text input by the user, or it can contain the original text input by the user and related dialogue context, thereby obtaining the target task that matches the user's intent more accurately based on the original text and related dialogue context.

[0051] The task splitter performs semantic understanding and logical reasoning based on the task description information. Specifically, it can identify multiple subdivided task requirements within the target task and the dependencies between these requirements. Based on the identification results, the task splitter divides the target task into multiple subtasks corresponding to the various task requirements and generates a corresponding capability requirement description for each subtask. In fact, the dependencies between subtasks are also simultaneously labeled, including sequential dependencies, parallel dependencies, and fan-out points. Sequential dependencies indicate which subtasks must wait for their predecessors to complete before starting; parallel dependencies indicate which subtasks can be executed simultaneously without affecting each other; and fan-out points indicate which nodes can dynamically expand to create more parallel execution branches.

[0052] For example, in a product analysis scenario, the objective is to "analyze the marketing strategy of a social platform product A over the past three months." The task decomposer breaks this objective down into the following sub-tasks: Sub-task 1 is "collecting publicly available data from product A over the past three months," with the following capability requirements: a data acquisition analyst responsible for acquiring and cleaning publicly available data and outputting a structured dataset; Sub-task 2 is "analyzing the user profile characteristics of product A," with the following capability requirements: a user profile analyst responsible for statistically summarizing and generalizing user tags based on data sources and outputting a user profile analysis report; Sub-task 3 is "analyzing product A's search term targeting strategy," with the following capability requirements: a search strategy analyst responsible for analyzing targeting channels and word class distribution related to search terms and outputting a search term analysis report. In practical applications, the task decomposer can indicate that sub-tasks 2 and 3 are parallel processes, both requiring sub-task 1 to be completed before starting.

[0053] The data processing method provided in the embodiments of this specification decomposes the target task and generates capability requirement descriptions corresponding to each sub-task. It transforms the vague and open natural language task description into a structured, executable set of sub-tasks with clear capability orientations. This avoids the decline in reasoning quality caused by unclear responsibility boundaries when a single intelligent agent processes the global task, and can improve the parallelism and overall efficiency of subsequent multi-agent collaboration.

[0054] In one or more embodiments of this specification, the user inputs a task to be processed. An admission-based intelligent agent assesses the task to determine whether it is a complex task (i.e., a target task of a specific type). Only if the task is complex is the multi-agent system used to process it. Specific implementation methods are described below: Before breaking down the target task into its components and generating at least one subtask and its corresponding capability requirements, the process also includes: Receive tasks to be processed and use an admission judgment agent to determine the task type of the tasks to be processed; If the task type of the task to be processed is a target type, the task to be processed is determined as a target task, wherein the target type is the type of task that needs to be processed by multiple agents.

[0055] In this context, the task to be processed can be understood as the original task request directly input by the user, and its form includes, but is not limited to, natural language text, text instructions converted from speech, or form data. The admission judgment agent can be understood as an agent specifically responsible for task type identification, used to classify and judge the input task. The task type can be understood as the category used to determine whether the task to be processed is suitable for multi-agent collaborative processing; this category is output by the admission judgment agent after semantic understanding and rule judgment of the task to be processed. The target type can be understood as a pre-defined type indicating that the task needs to be processed collaboratively by a multi-agent system, and its judgment criteria are as follows.

[0056] In the embodiments of this specification, the input of the admission judgment agent is the task to be processed, and the output is a binary result (allow / block) and a judgment reason. If the output result is allowed, it means that the task type of the task to be processed is the target type, and the multi-agent system is used to process the task to be processed.

[0057] When outputting a blocking label (i.e., non-target type) for tasks with severely incomplete information or that are clearly not team collaboration tasks (such as casual conversation), in the event that the admission judgment agent fails to judge (e.g., the confidence of the output result is lower than the preset threshold, or the admission judgment agent fails), in order to avoid blocking the processing of normal tasks due to judgment failure, a failure-open strategy will be adopted, that is, allowing the task to be processed to enter the processing flow of the multi-agent system as the target task.

[0058] In practical applications, the admission judgment agent adopts an invisible session mechanism in its judgment process. This invisible session mechanism can be understood as the system creating a one-time session with a specific prefix for each admission judgment. This session is invisible in the system session list, does not trigger subsequent processing, is not written to the persistent session history, and does not generate diagnostic logs. It is automatically destroyed after the judgment is completed, thereby achieving zero pollution of the session history.

[0059] Specifically, the system receives pending tasks input by the user through the client. Before sending the task to the multi-agent processing, it first inputs it to the admission judgment agent for type determination. After the admission judgment agent is invoked, it executes the judgment logic in an implicit session. This judgment logic is based on the personality definition of the admission judgment agent and the judgment rules preset in the task instruction file.

[0060] In practice, the admission judgment agent performs semantic analysis on the task to be processed to determine whether it belongs to a task with seriously incomplete information or obviously non-cooperative tasks. If it belongs to the above situations, it outputs a blocking label, indicating that the task to be processed does not belong to the target type, and returns the task to be processed by the single agent mode.

[0061] For example, in a social media data analytics platform application scenario, a user inputs "Analyze the market strategy of a social media platform product A over the past three months." After receiving this task, the system passes it to the admission assessment agent. The agent performs semantic analysis on the task description during the implicit session to determine its information completeness. The agent determines that the task is not severely incomplete and falls under the category of data analysis tasks, requiring complex processing through multi-agent collaboration. Therefore, the admission assessment agent outputs an allow label, classifying the task as the target task.

[0062] In another application scenario, a user inputs "What will the weather be like tomorrow?" into the system. The admission-determining agent judges that the task to be processed belongs to the casual conversation category, so it outputs a blocking label, returns the task to single-agent mode, directly generates a reply, and does not trigger the multi-agent collaboration process.

[0063] The data processing method provided in the embodiments of this specification achieves task splitting between multi-agent collaborative processing mode and single-agent processing mode by introducing a judgment based on admission judgment agent before task decomposition, thereby avoiding inefficient occupation of multi-agent system resources by non-collaborative tasks.

[0064] In one or more embodiments of this specification, when it is determined that a multi-agent collaboration mode will be used to process the target task, the existence of a predefined agent system matching the target task can first be determined from the already fixed set of predefined agent systems. If such a system exists, the target task can be directly processed using the predefined agent system. Otherwise, the target task is broken down, and the task execution agent corresponding to each sub-task is determined from the agent set, thereby constructing the multi-agent system corresponding to the target task. Specific implementation methods are described below: Before breaking down the target task, the process also includes: Based on the task description information of the target task, determine whether there is a predefined intelligent agent system that matches the target task; If it exists, the target task is processed using the predefined intelligent agent system; If it does not exist, perform the steps described above for task decomposition of the target task.

[0065] The task description information can be understood as textual information describing the content, objectives, and constraints of the target task. A predefined agent system can be understood as a pre-configured multi-agent team with predefined agents and collaborative processes between them. The agent composition, roles, functions, communication methods, and collaboration specifications of this predefined agent system are all predefined and fixed in configuration files.

[0066] In practical applications, predefined agent systems are suitable for handling tasks with high frequency, relatively fixed business models, and standardized processes. They can directly execute task processing without dynamically creating agents. Each predefined agent system is used to handle different task scenarios or types.

[0067] Specifically, the task description information of the target task is obtained, and based on this task description information, a suitable predefined intelligent agent system is matched in the predefined intelligent agent system set. In fact, each predefined intelligent agent system in the predefined intelligent agent system set may have corresponding matching rules or system description information of the task scenario. The task description information of the target task is compared with the system description information of each predefined intelligent agent system, and the matching degree is calculated.

[0068] If a predefined agent system exists that matches a preset threshold, it means that the set of predefined agent systems contains a fixed team configuration for this type of task. In this case, the predefined agent system can be directly invoked to process the target task. The target task is executed within the predefined agent system based on multiple predefined agents and collaborative processes, and the processing result is produced.

[0069] If there is no predefined agent system whose matching degree meets the preset threshold, it means that there is no fixed team configuration for this type of task in the set of predefined agent systems. At this time, the link of dynamically building a multi-agent system is entered, and the task decomposition step is performed for the target task, and the corresponding multi-agent system is dynamically built for the target task.

[0070] For example, in a product analytics scenario, a brand submits a task to the system every Monday: "Analyze the price trend of product A on a social platform over the past week." After the admission-based intelligent agent assesses the user's submitted task, it identifies the task as the target task and obtains the task description information. Based on this task description information, a matching process is then performed within a predefined set of intelligent agents.

[0071] Since this target task occurs on a weekly schedule and has a fixed task pattern (fixed analysis object, fixed analysis dimension, fixed time period), the multi-agent system for "product weekly report price analysis" can be configured as a predefined agent system. This predefined agent system is configured with multiple agents such as data acquisition agent, price analysis agent, and report generation agent. The system description information corresponding to the predefined agent system can be "applicable to the specified product price analysis task executed on a weekly schedule".

[0072] Therefore, by semantically matching the task description information of the above target task with the system description information, and if the matching degree is higher than the preset threshold, the "Product Weekly Report Price Analysis" is directly invoked to process the target task. During the execution process, no new intelligent agent needs to be dynamically created, which significantly shortens the task processing cycle.

[0073] The data processing method provided in the embodiments of this specification can directly reuse the fixed multi-agent team configuration for high-frequency repetitive tasks based on the matching between the target task and the predefined intelligent agent system. This avoids the consumption of computing resources and time overhead caused by repeatedly performing task decomposition, capability matching and dynamic creation of intelligent agents for fixed process tasks.

[0074] Step 204: Based on the capability requirements of each subtask, determine the corresponding task execution agent for each subtask from the set of agents.

[0075] In this context, the agent set can be understood as the collection of all available agents. In the following embodiments, the agent set can be referred to as the agent pool, which includes a list of all registered agents. In practical applications, this agent set includes pre-configured agents as well as historical agents that have been validated and added to the agent set. Each agent in the agent set has an independent workspace, model configuration, tool permissions, personality definition, and task instruction configuration. The task execution agent can be understood as an agent that executes specific sub-tasks.

[0076] Specifically, after breaking down the target task, for each subtask, the capability requirements of the subtask are matched with the agents in the agent set. In this embodiment, the agent with the highest matching degree with the subtask and meeting the preset matching threshold can be used as the task execution agent corresponding to the subtask. If a subtask cannot be matched with an agent that meets the preset matching threshold among all agents, the subtask can be marked as a target subtask that requires the creation of a new agent, thereby triggering the dynamic agent creation process.

[0077] In one or more embodiments of this specification, each intelligent agent in the intelligent agent set has corresponding intelligent agent configuration information. This allows for semantic matching based on the capability requirements information corresponding to a subtask and the intelligent agent configuration information, thereby determining a matching task execution intelligent agent for the subtask. Specific implementation methods are described below: Based on the capability requirements of each subtask, a corresponding task execution agent is determined from the set of agents, including: Determine the agent configuration information of each agent in the agent set, wherein the agent configuration information includes personality configuration information and task instruction configuration information; Semantically match the capability requirements of each subtask with the agent configuration information of each agent, and determine the matching agent for each subtask based on the semantic matching results.

[0078] Among them, the agent configuration information can be understood as a set of data describing the agent's ability characteristics, including personality configuration information and task instruction configuration information. Personality configuration information, namely personality definition, is used to define the agent's role identity, cultural background and behavioral perspective. Task instruction configuration information can be understood as a configuration file that defines the agent's task execution instructions and behavioral norms, which is used to specify how the agent should operate, how to call tools and how to organize output results when it receives a specific task.

[0079] Semantic matching can be understood as the process of calculating the similarity between the semantic representation of capability requirement information and the semantic representation of each agent's configuration information. This is used to select the agent from the set that best matches the capability requirements of the sub-task. The semantic matching process can be implemented based on the calculation of similarity between semantic vectors or based on keyword matching; neither is limited here. The semantic matching result can be understood as a similarity score or matching degree value between the capability requirement information and the agent's configuration information, used to quantify the degree of matching between the capabilities required for the sub-task and the actual capabilities of the agent.

[0080] In this embodiment, the capability requirement information corresponding to the subtask can be converted into a semantic vector representation, and the configuration information of each agent can also be converted into a semantic vector representation. Then, the similarity score between the two can be calculated. The similarity score is used to measure the degree of matching between the actual capability of the agent and the capability required by the subtask. Based on the calculated similarity score, each agent can be sorted or filtered to determine the corresponding task execution agent for the subtask.

[0081] For example, the subtask is "Analyze the acceptance of product appearance design based on collected user review data," and its role description in the capability requirements information is "Natural Language Processing Analyst." The system reads the configuration information of agent C, whose personality configuration information is "You are a social data analysis expert, skilled in comment sentiment analysis and user emotion recognition," and task instruction configuration information is "When receiving a text analysis task, call the sentiment analysis tool to determine the sentiment polarity of the text and output the statistical distribution." The descriptions of "sentiment analysis" and "emotion recognition" in agent C's configuration information highly match the "analysis acceptance" and "user reviews" of this subtask, so the system determines it as the executing agent for this subtask.

[0082] The data processing method provided in the embodiments of this specification can efficiently and accurately match the most suitable task execution agent for a subtask by calculating the semantic matching result between the capability requirement information corresponding to the subtask and the configuration information corresponding to the agent, without the need for manual annotation, thereby improving the automation level and matching accuracy of multi-agent system construction.

[0083] In one or more embodiments of this specification, if the semantic matching result between the capability requirement information corresponding to a subtask and the agent configuration information of any agent does not meet a preset threshold, then the subtask can be marked as a target subtask. A new agent needs to be constructed for the target subtask to respond to new requirements that exceed the coverage of the agent's capability pool. Specific implementation methods are as follows: Based on the semantic matching results, the matching task execution agents are determined for each subtask, including: If the semantic matching result between the capability requirement information corresponding to the subtask and the agent configuration information of any agent in the agent set meets a preset threshold, the corresponding agent is determined as the task execution agent corresponding to the subtask. Otherwise, mark the subtask as the target subtask that requires the creation of a new agent.

[0084] The preset threshold can be understood as a pre-set matching degree threshold, used to determine whether the semantic matching result has reached the minimum acceptable matching level.

[0085] Specifically, after calculating the semantic similarity between the capability requirement information and the configuration information of each agent, the semantic matching results between the subtask and each agent in the agent set are obtained. Each semantic matching result is then compared with a preset threshold.

[0086] If, for a given subtask, there exists an agent whose semantic matching result is greater than or equal to a preset threshold, then one agent is selected from these agents as the task execution agent for that subtask. In practice, the agent with the highest semantic matching result can be selected as the task execution agent, or an agent can be randomly selected from multiple agents whose semantic matching results exceed the preset threshold. This means that the corresponding task execution agent can be directly reused in subsequent executions of the subtask without triggering a creation process.

[0087] If, for a given subtask, the semantic matching results with each agent are found to be below a preset threshold, it indicates that there are no existing agents in the agent set that can effectively cover the capability requirements of the subtask. The subtask is then marked as a target subtask for which a new agent needs to be created. Specifically, a status identifier can be attached to the subtask, which is used to trigger the dynamic creation process of the agent in subsequent processes.

[0088] For example, when the target task is broken down into subtasks A, B, and C, the capability requirements of each subtask are semantically matched with the agent configuration information of agents A (search engine data expert), B (user profile analyst), and C (sentiment analysis expert) in the agent set to obtain the corresponding semantic matching results.

[0089] If the preset threshold is 0.70, for subtask A, agent A has the highest score of 0.93, meaning the semantic matching result between subtask A and agent A is greater than 0.70. Therefore, agent A can be identified as the task execution agent for subtask A. For subtask B, agent B has the highest score of 0.87, meaning the semantic matching result between subtask B and agent B is greater than 0.70. Therefore, agent B can be identified as the task execution agent for subtask B. For subtask C, the semantic matching results with agents A, B, and C are 0.56, 0.43, and 0.69 respectively. Since the semantic matching results between subtask C and each agent are below the preset threshold, subtask C is identified as the target subtask.

[0090] The data processing method provided in the embodiments of this specification, when it is determined that the subtask and the agent are successfully matched according to the semantic matching result, directly reuses the existing agent as the task execution agent of the subtask. When it is determined that the subtask is not matched according to the semantic matching result, a new process is triggered to ensure the integrity of task coverage. By using a preset threshold, the decision of whether to reuse the agent is transformed into a quantifiable comparison problem, thereby realizing automated and deterministic process determination.

[0091] Step 206: If there is no task execution agent corresponding to the target subtask in the set of agents, create a task execution agent corresponding to the target subtask based on the initial agent and the capability requirement information of the target subtask.

[0092] The initial intelligent agent can be understood as a basic intelligent agent that provides a built-in communication protocol. It includes basic capabilities such as model configuration, communication skeleton and workspace, but does not include the behavioral definition of specific business roles. The initial intelligent agent does not have the ability to execute specific tasks before human configuration information and task instruction configuration information are injected.

[0093] The general agent set can be understood as a collection of basic agents that can be used as execution shells. It should be noted that different general agents in the general agent set have differences in model capability dimensions. For example, some general agents excel in instruction compliance, some excel in complex reasoning, and some excel in the richness of language expression.

[0094] Specifically, if there is no task execution agent corresponding to the target subtask in the agent set, the agent creation process is triggered. In practice, the initial agent is reused from the general agent set as the execution shell, and role definitions, behavior specifications and special toolsets are injected into the initial agent according to the capability requirements of the target subtask, thereby dynamically creating a temporary task execution agent corresponding to the target subtask.

[0095] In the embodiments described in this specification, task execution agents that meet the capability requirements of new tasks can be generated on demand without relying on manual pre-definition, thereby expanding the response capability of multi-agent systems to new tasks.

[0096] In one or more embodiments of this specification, a temporary task execution agent corresponding to the target subtask is dynamically created by using a template injection method, with a general basic intelligent agent as the execution shell, and combining the role definition, behavior norms, and tool permissions in the capability requirement information corresponding to the target subtask. This allows for on-demand expansion of system capabilities without manual intervention. Specific implementation methods are described below: Based on the initial agent and the capability requirements of the target subtask, a task execution agent corresponding to the target subtask is created, including: Based on the capability requirement information of the target sub-task, an initial intelligent agent is determined from the set of general intelligent agents, wherein each general intelligent agent in the set of general intelligent agents has different processing capabilities in different processing dimensions; Based on the role information and task instruction information in the capability requirement information corresponding to the target sub-task, determine the personality configuration information and task instruction configuration information, and determine the role toolset corresponding to the role information; The personality configuration information and task instruction configuration information are injected into the initial agent, and the role toolset is configured for the initial agent to obtain the task execution agent corresponding to the target sub-task.

[0097] In this context, role information can be understood as descriptive information within the capability requirement information used to define the functional positioning required for sub-tasks. Personality configuration information can be understood as a configuration file defining the agent's role identity, cultural background, and behavioral perspective, used to assign the agent "who they are and what they are good at." Task instruction information can be understood as the required operations corresponding to sub-tasks within the capability requirement information. Task instruction configuration information can be understood as a configuration file defining the agent's task execution instructions and operational specifications, used to specify how the agent should operate after receiving a task, how to invoke tools, and how to organize the output results.

[0098] A role toolset can be understood as a set of tool permissions bound to a specific role. Tools in this tool permission set are only registered in the sessions of agents belonging to that role. Toolsets for different roles do not overlap, and tool permissions are gated by role. For example, scheduling agents have job assignment tools, execution agents have reporting tools, and communication coordination agents have broadcasting tools, thereby improving the accuracy of tool invocation.

[0099] In this embodiment, personality configuration information, task instruction configuration information, and role toolset are configured in the initial intelligent agent to realize the intelligent agent creation process. That is, the initial intelligent agent provides a fixed execution shell with process resources, communication skeleton, and workspace. Personality configuration information and task instruction configuration information are injected into business capability definitions, and role toolset is injected into operation permissions. The combination of the three generates a temporary task execution intelligent agent with complete execution capabilities.

[0100] Specifically, when the process of creating a new intelligent agent is triggered, the capability requirement information corresponding to the target subtask is obtained, the role description information (i.e., role information) is parsed from the capability requirement information, and the most suitable initial intelligent agent is selected from the general intelligent agent set according to the type characteristics of the role description information.

[0101] In fact, the general intelligent agent set contains multiple basic intelligent agents with different model capability characteristics. The system selects the initial intelligent agent according to the task characteristics involved in the target sub-task. For example, when the target sub-task requires high instruction compliance and role-playing ability, the system selects the initial intelligent agent with the corresponding model capability dimension (the initial intelligent agent has accurate instruction response and stable personality). When the target sub-task requires complex multi-step reasoning and deep logical analysis, the system selects the initial intelligent agent with the corresponding model capability dimension (the initial intelligent agent has a rigorous reasoning chain and accurate logic). When the target sub-task requires richness and diversity of language expression, the system selects the initial intelligent agent with the corresponding model capability dimension (the initial intelligent agent has a distinctive language style).

[0102] Given an initial intelligent agent, personality configuration information is generated based on the role information corresponding to the target sub-task. This means generating definition text for the initial intelligent agent's role identity, cultural background, and behavioral perspective. Additionally, task instruction configuration information is generated based on the task instruction information in the capability requirement information corresponding to the target sub-task. This means generating the execution instructions and behavioral norms corresponding to the initial intelligent agent.

[0103] In practice, the corresponding role toolset is determined based on the role information. The tool permissions in the role toolset correspond to the role's responsibility boundaries. For example, a role responsible for analysis has permission to read and analyze tools but not permission to assign tasks. After determining the above configuration information, the personality configuration information, task instruction configuration information, and role toolset are configured in the initial intelligent agent.

[0104] The personality configuration information and task instruction configuration information are used to define the business capabilities of the initial agent, while the tools in the role toolkit are used to define the operational permissions of the initial agent. After configuration, the initial agent is endowed with the ability to execute specific responsibilities for the target subtask, thereby generating a temporary task execution agent corresponding to that target subtask. It should be noted that the task execution agent corresponding to this target subtask is temporary and distinct from the fixed agents in the agent set.

[0105] For example, in a data analysis scenario, the target subtask is "analyzing user behavior logs and performing anomaly detection". The role information in its capability requirement information is "big data processing analyst". Based on this role information, a basic intelligent agent with large-scale data processing capabilities is selected from the general intelligent agent set as the initial intelligent agent, and the corresponding data query and anomaly detection toolset is configured to obtain the corresponding temporary task execution intelligent agent.

[0106] The data processing method provided in the embodiments of this specification configures personality configuration information, task instruction configuration information and role toolset for the initial intelligent agent by using the capability requirement information of the target sub-task and template injection. This enables the newly created intelligent agent to obtain complete business capability definition and operation permissions while reusing the communication skeleton of the basic intelligent agent, ensuring that the newly created intelligent agent does not deviate from the system's communication specifications.

[0107] Step 208: Execute at least one subtask using the scheduling agent and the task execution agent corresponding to each subtask to obtain the task processing result.

[0108] In this context, the scheduling agent can be understood as a centralized scheduling component in a multi-agent system responsible for task routing, dispatching, acceptance, and quality control. It is used to uniformly manage the execution progress, message communication, and result aggregation of each task instance. A task instance can be understood as an execution unit generated by binding subtasks and their corresponding task execution agents. This execution unit has an independent execution context, task responsibility description, and lifecycle state. One task execution agent can simultaneously host multiple task instances.

[0109] Specifically, given that the task execution agents corresponding to each subtask are determined and a corresponding task execution agent is created for the target subtask, a multi-agent system (multi-agent team) corresponding to the target task can be constructed based on each task execution agent and the scheduling agent. In this way, the multi-agent team can be used to process at least one subtask and finally obtain the task processing result corresponding to the target task.

[0110] In one or more embodiments of this specification, after matching or creating the task execution agents corresponding to each subtask, an independent task instance is created for each subtask and the task instance is bound to the task execution agent, so that each subtask obtains an independent execution context and responsibility boundary; the scheduling agent performs unified scheduling of each task instance according to the dependency relationship between subtasks, thereby realizing the orderly collaboration of a multi-agent team. Specific implementation methods are described below: Executing at least one subtask using a scheduling agent and a task execution agent corresponding to each subtask includes: Create corresponding task instances for each subtask, and bind each task instance to the corresponding task execution agent; Based on the dependencies between subtasks, the scheduling agent schedules each task instance.

[0111] In this context, a task instance can be understood as an execution clone created on top of a task execution agent, based on the specific responsibilities of a task. The same basic agent can simultaneously carry multiple task instances to execute different analytical aspects in parallel. For example, the same agent with data query capabilities can simultaneously carry a task instance of "querying user behavior logs" and a task instance of "querying search term statistics". The two task instances execute in parallel without interfering with each other.

[0112] Binding can be understood as associating and mapping the task responsibility description (a triplet structure defining the responsibility of the task instance, including role description, role boundaries, and delivery instructions) corresponding to the subtask with the execution capabilities of the task execution agent, so that the task instance can call the model configuration, tool permissions, and personality definition of the bound task execution agent when executing the subtask; in fact, the task instance itself does not independently own the execution resources of the corresponding task execution agent, and its execution capabilities come from the bound task execution agent.

[0113] The dependencies between subtasks can be understood as the constraints on the execution sequence of each subtask marked during the task splitting stage, including serial dependencies and parallel dependencies. Serial dependencies indicate that there is a production dependency between subtasks, that is, a subtask must wait for another subtask to complete and can only start based on the output of the previous subtask. Parallel dependencies indicate that there is no production dependency between subtasks and they can be executed simultaneously.

[0114] Specifically, each subtask is bound to its corresponding task execution agent, generating task instances for each subtask. The dependencies between subtasks are then obtained and mapped to the corresponding task instances. The scheduling agent can determine the execution order of task instances based on these dependencies. For example, task instances without prerequisite dependencies are marked as parallelizable; task instances with prerequisite dependencies are marked as waiting, and executed only after the prerequisite task instances they depend on have completed and produced results.

[0115] The data processing method provided in the embodiments of this specification creates independent task instances for each subtask, enabling each subtask to obtain an independent execution context and clear responsibility boundaries. By scheduling an intelligent agent to uniformly schedule each task instance according to the dependency relationship between subtasks, task instances with dependencies are executed serially according to the correct dependency relationship, ensuring the correctness of multi-agent collaboration.

[0116] In one or more embodiments of this specification, for the second type of subtasks that need to be executed serially, communication between the task instance and the scheduling agent is achieved through an inbox. This inbox information can be persistently stored in a storage medium and may include the sender, receiver, timestamp, and specific message content, facilitating subsequent backtracking. Specific implementation methods are described below: Based on the dependencies between subtasks, the scheduling agent schedules each task instance to execute its corresponding subtask, including: Based on the dependencies between subtasks, determine the first type of subtasks to be executed in parallel and the second type of subtasks to be executed sequentially; For the first type of subtask, the scheduling agent schedules the corresponding task instances in parallel for execution. For the second type of subtask, the scheduling agent schedules the corresponding task instances in sequence according to the dependency order for execution. The scheduling agent sends execution instructions to the task instance through the inbox corresponding to the task instance, and the task instance returns task execution information to the scheduling agent through the inbox corresponding to the scheduling agent.

[0117] The first type of subtask can be understood as a set of subtasks that have no output dependency relationship and can be executed simultaneously without affecting each other; the second type of subtask can be understood as a set of subtasks that have output dependency relationships and need to be executed in a specific order; in the second type of subtask, the subtask executed later depends on the output result of the preceding subtask as input, so it needs to wait for the preceding subtask to complete and produce a result before it can start.

[0118] The inbox can be understood as an asynchronous persistent communication channel between agents, supporting point-to-point sending and broadcasting messages. Inbox information can be persistently stored on disk so as not to be lost after the process restarts. The scheduling agent sends execution instructions to the task instance through the inbox, and the task instance sends the execution results back to the scheduling agent through the inbox.

[0119] It should be noted that the inbox corresponding to the task instance and the inbox corresponding to the scheduling agent can be understood as different instances of the same inbox system. The scheduling agent sends a message to the inbox of the target task instance, and the task instance reads and executes it; after the task instance completes, it sends a message to the inbox of the scheduling agent, and the scheduling agent reads and continues scheduling.

[0120] Specifically, for any two subtasks, if there is no direct serial dependency between them, and no indirect dependency after passing through dependency relationships, then they belong to the first category of subtasks that can be executed in parallel. For two subtasks that have a direct dependency or an indirect dependency after passing through dependency relationships, then they need to be executed serially and belong to the second category of subtasks. Following this classification method, all subtasks are divided into the first category of subtasks that can be executed in parallel and the second category of subtasks that need to be executed serially.

[0121] For multiple subtasks classified as subtasks of the first category, the scheduling agent sends execution instructions in parallel to the inboxes of the corresponding task instances during scheduling execution. Each task instance can start execution simultaneously after receiving the instruction, and the execution process does not wait for each other. After execution, each task instance sends the execution result back to the inbox of the scheduling agent.

[0122] For multiple subtasks classified as subtasks of the second category, the scheduling agent schedules them according to the dependency order determined by the dependency relationship. Specifically, the scheduling agent can send execution instructions to the inbox of the task instance with the highest dependency order. After execution, this task instance sends its execution result back to the scheduling agent's inbox. Upon receiving this execution result, the scheduling agent uses it as input data or a trigger condition for subsequent task instances and continues sending execution instructions to the inboxes of subsequent task instances according to the dependency order, until all task instances in the second category are executed sequentially.

[0123] Communication between the scheduling agent and each task instance is completed through the inbox. The persistent storage mechanism of the inbox ensures that messages are not lost after the process restarts, and task instances or scheduling agents can read messages in the inbox at any time for asynchronous processing.

[0124] The data processing method provided in the embodiments of this specification ensures data integrity by serially scheduling dependent subtasks in the order of dependency. Furthermore, when the scheduling agent and task instance complete the instruction issuance and result return through the inbox, asynchronous persistent communication between the scheduling agent and task instance can be realized, thereby improving the reliability of the multi-agent system.

[0125] In one or more embodiments of this specification, during the process of a task instance executing a subtask and sending the execution result back to the scheduling agent, the task instance may attach structured challenge information. This challenge information is used to request supplementary information from other task instances to improve its own analysis conclusions. The scheduling agent can forward this challenge information to the inbox of the target task instance, thereby realizing coordinated communication between task instances. Specific implementation methods are described below: If any task instance sends a task execution message to the inbox corresponding to the scheduling agent containing structured challenge information, the scheduling agent is used to determine the target task instance being challenged. The structured challenge information includes a target role identifier and follow-up questions. The target task instance is the task instance corresponding to the target role identifier. The scheduling agent sends follow-up questions to the target task instance through the inbox of the target task instance, and obtains the task execution information returned by the target task instance based on the follow-up questions.

[0126] In this context, task execution information can be understood as structured data returned by a task instance to the scheduling agent after completing its subtasks. This task execution information may include the execution results of the subtasks and optional, accompanying structured challenge information. Structured challenge information can be understood as a formatted information completion request proactively triggered by a task instance when it discovers that certain data or logic involved in the conclusions depend on the analytical perspectives of other roles.

[0127] In this embodiment, the structured interrogation information includes fields such as target role identifier and follow-up question content. The target role identifier describes the capabilities or role type required by the target task instance, while the follow-up question content describes specific information that needs to be supplemented or specific conclusions to be verified. The target role identifier can be understood as descriptive information about the role identity of the target task instance, which needs to be determined from the role identifiers of existing task instances in the multi-agent system. The target task instance can be understood as the task instance to which the interrogated party belongs, corresponding to the target role identifier. The follow-up question content can be understood as an executable and clearly targeted question description.

[0128] Specifically, after receiving task execution information sent by any task instance through the inbox, the scheduling agent first parses whether the task execution information contains structured challenge information. In practical applications, the data structure of the task execution information includes an optional challenge field. When a task instance finds during execution that its conclusion requires cross-validation from the perspective of other roles or data supplementation from other roles, structured challenge data is filled into this field.

[0129] In practice, when the task execution information contains cross signals, it is determined that the task execution information contains structured challenge information. The target role identifier and follow-up questions are then extracted from this structured challenge information. The scheduling agent matches the target role identifier with the role identifiers of each task instance in the current multi-agent team to determine the target task instance corresponding to that target role identifier. After determining the target task instance, the scheduling agent sends the follow-up questions to the target task instance through its inbox. Upon receiving the follow-up questions, the target task instance treats them as additional queries. For additional queries, after obtaining the response, it needs to return a challenge reply to the scheduling agent through its inbox.

[0130] After receiving the challenge response from the target task instance, the scheduling agent associates and stores the challenge response with the original challenge information to form a complete challenge-response record.

[0131] In the embodiments of this specification, when the structured challenge information also includes a priority field, the scheduling agent adopts a differentiated processing strategy based on the priority identifier. The priority identifier is used to indicate the strength of the downstream conclusion's dependence on the challenge content, including high priority and medium priority. High priority means that the downstream conclusion strongly depends on the response to the challenge content. Therefore, high priority challenge information is immediately routed and forwarded, while medium priority challenge information can be temporarily stored in the pending list. After all task instances are completed, the scheduling agent decides whether to follow up.

[0132] The routing rule for the scheduling agent regarding the challenge content is as follows: for high-priority challenge information, the agent matches the target role identifier in the challenge information with the role in the multi-agent team. If the match is successful, the agent sends follow-up questions to the target task instance through the inbox. If the match fails (e.g., the semantic similarity between the target role identifier and the personality configuration information of any agent in the multi-agent team is lower than the preset similarity threshold), the high-priority challenge information is downgraded to a medium-priority challenge information, and the challenge information is marked as unrouted.

[0133] In the questioning process, questioning behavior can be restricted through anti-recursion constraints and high-priority signal limits. Anti-recursion constraints prevent the target task instance's response to structured questioning information from triggering new structured questioning information. This means that even if the response contains new questioning information, it will not be recognized and processed by the scheduling agent. Furthermore, the same structured questioning information triggers only one question, preventing the same follow-up question from being repeatedly sent in multiple rounds of dialogue. High-priority signal limits constrain the total number of high-priority questioning messages; that is, the number of high-priority structured questioning messages that each task instance can generate in a single execution cannot exceed a preset limit (e.g., 3 messages).

[0134] By restricting the questioning behavior through anti-recursion constraints, the infinite loop of follow-up questions between multiple task instances is prevented, thus ensuring the termination and controllability of the questioning process.

[0135] The data processing method provided in the embodiments of this specification overcomes the problems of convergence uncertainty and attribution difficulties caused by the participation of all members in free group chat collaboration by scheduling intelligent agents to uniformly receive, parse and route structured question information, and by using an inbox to make the question process have traceable communication records. Furthermore, the matching mechanism between target role identifiers and task instances enables the question content to be accurately delivered to target task instances with corresponding capabilities and perspectives, avoiding the communication overhead caused by undirected broadcasting.

[0136] In one or more embodiments of this specification, after each task instance completes execution and returns the execution result to the scheduling agent, the execution result of each sub-task can be quality-checked to ensure that the output of each sub-task meets the preset quality standards. For execution results that fail the check, the scheduling agent sends a modification instruction to the corresponding task instance, which then resubmits the revised result for acceptance. After all task instances pass the acceptance, the scheduling agent integrates the execution results to form the final task processing result. Specific implementation methods are described below: Obtaining task processing results also includes: The execution results of each task instance are accepted according to the preset acceptance criteria to obtain the acceptance results of each task instance. If the acceptance result is unsuccessful, a modification instruction is sent to the corresponding task instance, the modified execution result is received, and the acceptance is performed again. After the execution results of each task instance have passed acceptance, the execution results are integrated to obtain the task processing result.

[0137] The preset acceptance criteria can be understood as a set of structured rules used to measure the output quality of each subtask, including measurable requirements for the completeness, accuracy, and format standardization of the execution results of each subtask. The acceptance result can be understood as the quality judgment result output by the scheduling agent after comparing the execution result of the task instance with the preset acceptance criteria, including two states: pass and fail.

[0138] A modification instruction can be understood as a structured revision request sent by the scheduling agent to a task instance that failed acceptance. This instruction may include an explanation of the reason for failure, specific modification requirements, etc. Integrating the execution results can be understood as the scheduling agent merging and organizing the execution results of each task instance according to a preset report template or summary logic after all subtasks have passed acceptance, forming a unified task processing result for the user.

[0139] Specifically, after each task instance completes execution and returns the execution results, the scheduling agent compares the execution results of each task instance item by item based on the corresponding preset acceptance criteria. The comparison content includes, but is not limited to, whether the execution results contain complete required fields, whether the data format meets the specification requirements, and whether the conclusions have traceable data basis, etc., which are not limited here.

[0140] For execution results that meet the preset acceptance criteria, the scheduling agent marks the acceptance status of the execution result as passed. For execution results that do not meet the preset acceptance criteria, the scheduling agent marks the acceptance status of the execution result as failed and generates a corresponding modification instruction. This modification instruction can clearly indicate the specific reason for the failure (e.g., "missing data source") and the specific content that needs to be modified. Then, the modification instruction is sent to the corresponding task instance through the inbox.

[0141] Upon receiving a modification instruction, the task instance revises its execution result according to the instruction and sends the revised result back to the scheduling agent. The scheduling agent then verifies the revised result again. If it still fails, the process of sending the modification instruction and receiving the revised result is repeated until the task instance's execution result passes verification.

[0142] After the execution results of each task instance pass acceptance, the scheduling agent integrates the execution results according to a preset aggregation logic to obtain the final task processing result. It should be noted that, given that all execution results have passed acceptance, an aggregation task instance (i.e., a task instance generated based on the aggregation subtask and the corresponding task execution agent) can be used to aggregate multiple execution results, thereby obtaining the integrated task processing result. If the aggregation task instance encounters contradictory conclusions between different execution results, the contradiction information (which may include the contradictory execution results, the corresponding task instance, etc.) is reported to the scheduling agent, allowing the scheduling agent to make a decision, rather than making a decision itself.

[0143] The data processing method provided in the embodiments of this specification automatically controls the quality of the execution results of task instances through preset acceptance criteria, so that execution results that do not meet the quality requirements can be detected and corrected before the final result is produced, avoiding the impact of low-quality or non-standard sub-task outputs on the accuracy of the final task processing result.

[0144] In one or more embodiments of this specification, during the execution of subtasks by multiple agents, to allow users to understand the task progress in real time without actively asking, translated natural language information can be sent by a dedicated communication coordination agent that sends messages to the user client. Specific implementation methods are described below: Executing at least one subtask using a scheduling agent and task execution agents corresponding to each subtask to obtain task processing results, further includes: The scheduling agent and the task execution agent corresponding to each subtask are used to execute the at least one subtask. During the execution, the task execution information received by the scheduling agent is copied to the communication coordination agent. Once the communication coordination agent determines that the task execution information meets the broadcasting conditions, it converts the task execution information into natural language information and outputs it to the corresponding user client.

[0145] In this context, "copy and paste" can be understood as automatically sending a complete copy of a message to the inbox of the communication coordination agent when a task instance sends the message to the scheduling agent, thus enabling the communication coordination agent to obtain an information view equivalent to that of the scheduling agent. Copy and paste is a one-way push, and the communication coordination agent receiving the copy does not affect the scheduling agent's normal receipt of the original, with zero change in the scheduling process.

[0146] The communication coordination agent can be understood as an agent that is independent of the scheduling agent and the task execution agent and is specifically used for user interaction. As the only information outlet in the multi-agent system, it translates the internal collaboration progress into user-readable natural language. In this embodiment, when creating the multi-agent system corresponding to the target task, the communication coordination agent is treated as an agent member. This communication coordination agent has an independent session context, an independent workspace, and a dedicated skill module.

[0147] The broadcasting conditions can be understood as the basis for the communication coordination agent to determine whether a task execution information is worth broadcasting to the user. In the embodiments of this specification, when the task execution information belongs to the type of opening confirmation, progress notification, final delivery report or user decision request, the task execution information is confirmed to meet the broadcasting conditions, while messages without substantial progress content or no information increment do not meet the broadcasting conditions, and the communication coordination agent returns a non-broadcasting flag for such messages.

[0148] Natural language information can be understood as the transformation of technical or semi-structured information for intelligent agents into fluent and easy-to-understand natural language expressions for human users. This transformation process usually involves removing internal protocol tags and converting technical terms into business language that users can understand.

[0149] Specifically, during the execution of at least one subtask using the scheduling agent and the task execution agents corresponding to each subtask, when the central scheduler delivers inbox messages to the scheduling agent, the system automatically copies a complete copy of each message to the inbox of the communication coordination agent.

[0150] In practice, the inbox message contains the sender, the recipient, and specific task execution information. Therefore, during the copy process, the source of the task execution information can be determined based on the sender in the inbox message. If the task execution information comes from the communication coordination agent itself, the copy is skipped to avoid message looping; if the message source is another task execution agent, the copy operation is performed.

[0151] During the copy operation, the team affiliation of each agent can be determined based on the session registry. Therefore, the task execution agent that sends the task execution information to the scheduling agent and the corresponding multi-agent team (i.e., the agent system) can be identified from the session registry. The communication coordinating agent corresponding to the multi-agent team is searched for in the session registry by role identifier. If found, a copy of the task execution information is sent to the inbox of the communication coordinating agent. If not found, the copy operation can be skipped to avoid blocking the scheduling process.

[0152] When a communication coordination agent is triggered by a message arrival event, it assesses the task execution information. If the information meets the broadcasting conditions, the agent converts it into natural language. This conversion process includes removing internal protocol markers, retaining key data, and reorganizing the information in business language. After conversion, the agent uses a broadcasting tool to output the natural language information to the user's corresponding client. If the broadcasting conditions are not met, the agent returns a no-broadcast flag, and no information is output.

[0153] In practical applications, given the defined task execution agents for each sub-task, a multi-agent system is constructed based on the calling agent, the task execution agent, and the communication coordination agent. The communication coordination agent is identified as "communicator," serving as a unique identifier for this role in the session registry. This communication coordination agent can be built using a basic agent with text organization capabilities as its execution shell. Furthermore, the task responsibility description of the communication coordination agent includes a role description ("team communicator, translating internal team progress into language users can understand"), role boundaries ("only extracts information without re-analysis, does not generate business documents, does not participate in scheduling, does not repeatedly broadcast the same progress, and returns NO_REPLY for messages without broadcast value"), and delivery instructions ("receives an event → determines whether it is worth informing the user → broadcasts if worthwhile, remains silent if not"). By defining the task responsibility description of the communication coordination agent, it prevents the agent from arbitrarily changing the task execution information to be output, instead fulfilling its function of broadcasting output.

[0154] The data processing method provided in the embodiments of this specification automatically copies the task execution information received by the scheduling agent to the communication coordination agent during the execution process, so that the communication coordination agent can obtain an information view equivalent to that of the scheduling agent. The communication coordination agent is used to judge the broadcast conditions and convert the task execution information into natural language, so that the user can obtain clear and understandable real-time progress reports in the collaborative execution of multiple agents.

[0155] In one or more embodiments of this specification, during the process of converting task execution information into natural language information, the communication coordination agent can organize the language according to the target output style, so that the broadcast content maintains a consistent style in terms of tone, wording, and information presentation. Furthermore, it needs to verify the converted, output natural language information to detect whether any technical markers related to internal communication of the agent remain, and replace them with user-readable natural language expressions when technical markers are detected. Specific implementation methods are described below: Converting the task execution information into natural language information and outputting it to the corresponding user client includes: Based on the target output style corresponding to the communication coordination agent, the task execution information is converted into natural language information with the target output style; Verify whether the natural language information contains internal technical tags; if it does, replace the internal technical tags with natural language information in the target output style. The communication coordination agent invokes a broadcasting tool to output verified natural language information to the user's corresponding client.

[0156] The target output style can be understood as the language style characteristics adopted by the communication coordination agent when broadcasting to users. It is defined in the personality definition information or exclusive skill module of the communication coordination agent, including but not limited to tone and expression preferences, which are not limited here. Internal technical tags can be understood as protocol tags and structured fields used in collaborative communication within a multi-agent system, including but not limited to status tags, task identifiers, and session identifiers. These tags have clear semantics for communication between agents, but are not understandable when facing users.

[0157] Verification can be understood as the process of scanning and detecting the content of natural language information to identify whether any internal technical markers remain. A broadcasting tool can be understood as a tool specifically designed to send messages to user clients. This broadcasting tool is the only legitimate way to send messages to users. Therefore, if this broadcasting tool is only registered in the session of the communication coordinating agent, and not in the sessions of other roles (such as the scheduling agent or task execution agent), it ensures that the only outlet for messages is the communication coordinating agent. The scheduling agent and task execution agent, lacking a broadcasting tool, cannot directly speak to the user.

[0158] In practical applications, the whitelist of tools corresponding to the communication coordination agent includes reading tools, broadcasting tools, and reporting tools (used to report the task loop to the scheduling agent when the output is completed). The communication coordination agent is restricted from having writing tools, execution tools, job assignment tools, or session management tools, so that the communication coordination agent does not participate in business analysis, job assignment, file writing, or session management.

[0159] Specifically, when the communication coordination agent determines that the task execution information meets the broadcasting conditions, it determines the target output style corresponding to the communication coordination agent. This target output style can be determined by the style parameters in the currently loaded exclusive skill module or personality definition information of the communication coordination agent. Different skill modules can define different output styles. Therefore, by loading different exclusive skill modules or adjusting the style parameters in the personality definition information, the output style can be plugged in and replaced. Moreover, the update operation does not require modification of the configuration information of the scheduling agent or the task execution agent, so it will not affect other agents in the multi-agent system.

[0160] It should be noted that the dedicated skill module corresponding to the communication coordination agent also defines the writing specifications for broadcasts, such as the following: Authenticity guarantee: Numbers, percentages, and time periods must faithfully represent the original materials; no fabrication, modification, or additions are permitted. Expression richness: Encourage the use of emojis, priority tags, tables, etc. to enhance readability; Product path reference: Provide the absolute path if a file exists, and explain the situation if no file exists; Closing the question: Provide specific, answerable follow-up questions based on the analysis, rather than clichés.

[0161] The communication coordination agent organizes and transforms the task execution information into language based on the target output style, generating natural language information with that style characteristics. Once the transformation is complete, the communication coordination agent verifies the natural language information, checking for internal technical markers, such as status markers, task identifier prefixes, session identifier formats, and other protocol fields used in internal communication.

[0162] If no internal technical markers are detected, the natural language information proceeds directly to the broadcasting stage. If internal technical markers are detected, the communication coordination agent replaces the detected technical markers with user-readable natural language expressions. The verified natural language information contains no internal technical markers and consists entirely of user-readable natural language content. The communication coordination agent then invokes the broadcasting tool to output the verified natural language information to the user's corresponding client.

[0163] The data processing method provided in the embodiments of this specification organizes language according to the target output style through communication coordination agents. By decoupling the output style from the scheduling logic in the multi-agent system, the independent configuration and flexible replacement of the output style are realized. Through a verification mechanism, internal technical tags in the natural language information are detected and replaced to ensure that no protocol tags or structured fields for internal communication of the agents remain in the information output to the user, so that the user can obtain understandable natural language broadcasts and avoid the negative impact of internal jargon on the user experience.

[0164] In one or more embodiments of this specification, when outputting natural language information to a user client, the system filters the message source through a whitelist mechanism, allowing messages originating from the communication coordination agent to be displayed to the user, thus preventing messages from the scheduling agent or task instance from bypassing the communication coordination agent and being directly exposed to the user. Specific implementation methods are described below: Outputting a verified natural language message to the user's corresponding client, including: When the natural language information originates from the communication coordination agent, the natural language information is output to the user's corresponding client.

[0165] In the embodiments described in this specification, the existence of a communication coordination agent in the multi-agent system is determined by the session registry. If such an agent exists, a whitelist mode is activated.

[0166] In whitelist mode, the session display layer will only allow messages originating from the communication coordination agent to be displayed to the user client for messages sent by the multi-agent system (including the scheduling agent, task execution agent, and communication coordination agent).

[0167] For example, if a data acquisition task instance sends the execution result back to the scheduling agent after completing data acquisition, and if the message is incorrectly attempted to be output to the user client, the display layer obtains the source marker of the message as "data_collector", which does not match the identity identifier "communicator" of the communication coordinating agent in the session registry. The message is then filtered by the display whitelist. However, if the communication coordinating agent announces "data acquisition complete", the display layer determines that the source marker of the message is "communicator", which matches the whitelist, and therefore it can be allowed to be sent to the user client.

[0168] In practice, if the session registry determines that there is no communication coordinating agent in the multi-agent system, a blacklist mode is enabled to ensure compatibility with older session versions. By switching between whitelist and blacklist modes based on the presence or absence of a communication coordinating agent with the multi-agent system, compatibility with historical sessions is guaranteed.

[0169] The data processing method provided in the embodiments of this specification uses a whitelist mechanism to filter message sources in the display layer, allowing only messages from the communication coordination agent to be displayed to the user, effectively preventing the accidental exposure of internal messages from non-communication coordination agents to the user.

[0170] In one or more embodiments of this specification, when the communication coordination agent determines whether task execution information needs to be broadcast to the user, it performs classification matching based on predefined broadcast types. Information belonging to the preset type is determined to meet the broadcast conditions, while information not belonging to the preset type is not broadcast, thereby providing a deterministic classification basis for broadcast determination. Specific implementation methods are as follows: Determine if the task execution information meets the broadcast conditions, including: If the communication coordination agent determines that the task execution information belongs to a predefined broadcast type, it determines that the task execution information meets the broadcast conditions.

[0171] The predefined broadcast types can be understood as pre-defined message categories that need to be broadcast to users, including opening confirmation, progress report, final delivery report, and user decision-making requirements. Each broadcast type corresponds to a specific trigger scenario and output content requirements. For example, the trigger scenario for the opening confirmation type is the startup message after the multi-agent team is successfully created, and the output content requirements include acknowledging receipt of the task and briefly describing the work to be carried out. The trigger scenario for the progress report type is the copy event corresponding to the execution result sent to the scheduling agent after the task instance completes the sub-task, and the output content requirements are to describe the current progress and the next plan in 2-4 sentences. The trigger scenario for the final delivery report type is the delivery of delivery materials after the scheduling agent receives the execution results of each task instance, and the output content requirements are to output a structured delivery report containing the product path and follow-up information. The trigger scenario for the user decision-making requirements type is the scheduling agent submitting a decision request, and the output content includes the missing decision, the optional decision, and the impact of each decision.

[0172] Specifically, after receiving the copied task execution information, the communication coordination agent extracts the content features of the information and matches them against categories in a predefined set of broadcast types. If the type field in the task execution information corresponds to one of the following types: opening confirmation, progress notification, final delivery report, or user decision-making requirement, or if the semantic content of the task execution information matches the features of any of the above types, the communication coordination agent determines that the task execution information meets the broadcast conditions. If the task execution information cannot be matched with any predefined broadcast type, it is determined that the information does not meet the broadcast conditions.

[0173] The data processing method provided in the embodiments of this specification classifies and determines task execution information through predefined broadcast types, so that the broadcast condition judgment has clear and executable classification basis, avoiding the uncertainty caused by relying on the model to make subjective judgments on "whether it is worth broadcasting", and realizing the deterministic triggering of broadcast timing.

[0174] In one or more embodiments of this specification, after obtaining the task processing result, the quality of dynamically created temporary intelligent agents is evaluated based on the acceptance results of each task instance. This allows qualified temporary intelligent agents to be solidified into permanent intelligent agents and stored in the intelligent agent set. Specific implementation methods are described below: After obtaining the task processing results, the following is also included: Based on the acceptance records of the execution results corresponding to each task instance, determine the quality level of the task execution agent corresponding to each target subtask; If the quality level is qualified, the task execution agent corresponding to the target subtask is solidified as the target execution agent and stored in the agent set.

[0175] Here, quality level can be understood as the grade label output by the corresponding task execution agent based on the acceptance record of the task instance, including qualified grade, observable grade, and poor grade; solidification can be understood as the process of persistently storing the temporarily created task execution agent from a temporary state to a permanent configuration, and the solidified agent is still available after the system restarts. Target execution agent can be understood as a permanent agent that has been solidified and registered in the agent set.

[0176] Specifically, after completing the acceptance of all task instances, the scheduling agent obtains the acceptance record of the task instance corresponding to each task execution agent marked as temporary, and determines the quality level of the task execution agent based on the acceptance record.

[0177] For example, an agent that passes acceptance on the first attempt is marked as "qualified"; an agent that passes acceptance after one revision is marked as "observational"; and an agent requiring review, exceeding the time limit, or needing to be reassigned is marked as "poor". For temporary agents with a quality level of "qualified", the system saves their personality configuration information and task instruction configuration information from the temporary workspace to the persistent storage workspace, and registers the agent in a separate persistent archive file so that the agents in the archive file can be automatically merged into the agent set when the system starts. The saved agents can be reused in subsequent tasks without being recreated.

[0178] By solidifying qualified, temporary task execution agents, a task closed loop is achieved. When a target task arrives, temporary agents can be dynamically generated. These temporary agents execute corresponding sub-tasks to obtain execution results. If the execution results pass the first acceptance test, the quality of the temporary agents is determined to be qualified. The temporary agents are then solidified into reusable permanent agents and stored in an agent set. This expands the task capability range covered by the agent set, enabling the reuse of agents when processing similar tasks corresponding to the target task.

[0179] The data processing method provided in the embodiments of this specification classifies temporary intelligent agents based on acceptance results and solidifies qualified temporary intelligent agents into permanent intelligent agents and stores them in the intelligent agent set. This enables the system to form a closed loop of capability growth of "dynamic creation - execution - acceptance - accumulation - reuse". As the number of task executions increases, reusable intelligent agents are continuously accumulated. Subsequent similar tasks can skip the dynamic creation process and directly reuse existing intelligent agents, which reduces system overhead and improves task execution efficiency.

[0180] In one or more embodiments of this specification, after obtaining the task processing result, the execution session records, inbox messages, and execution results of each task instance can be persistently saved to form a traceable chain of evidence. When a user questions a conclusion in the final report and initiates a tracing request, the system can trace back the task instance and basis for that conclusion based on the chain of evidence, thereby achieving the interpretability and verifiability of the conclusion. Specific implementation methods are as follows: After obtaining the task processing results, the following is also included: Save the execution session records, inbox messages, and execution results of each task instance to form a chain of evidence; In response to a request to trace the source of the target conclusion in the task processing result, the task instance that produced the target conclusion and the conclusion basis corresponding to the target conclusion are determined according to the chain of evidence.

[0181] The execution session record can be understood as a complete dialogue or reasoning record of the task instance during the execution of subtasks, including instructions received by the task instance, tools invoked, and intermediate reasoning steps. The inbox messages can be understood as all message records exchanged between the scheduling agent and the task instance through the inbox, including the sender, receiver, timestamp, and message body, persistently stored on disk. The execution result can be understood as the final conclusion or structured data produced by the task instance.

[0182] A chain of evidence can be understood as a complete traceability path from the final conclusion to the original data, consisting of execution session records, inbox messages, and execution results. A tracing request can be understood as a user's questioning or challenge of a specific conclusion in the task processing results.

[0183] Specifically, the execution session records of each task instance, the inbox messages between the scheduling agent and the task instance, and the execution results produced by each task instance are persistently stored on the disk. The above three types of data are associated and indexed according to the task instance identifier, timestamp, and message chain to form a traceable evidence chain.

[0184] When a user requests a source for a target conclusion in the task processing results, the description information or location identifier of the target conclusion is extracted from the request and searched in the evidence chain. Specifically, the task instance that produced the target conclusion can be determined by matching the content in the execution results of each task instance, the conclusion reporting records in each inbox message, and the reasoning process in each execution session record. The original data or intermediate results on which the task instance based its conclusion can also be located.

[0185] The data processing method provided in the embodiments of this specification forms an evidence chain by persistently saving execution session records, inbox messages and execution results, so that every conclusion in the final report can be traced back to the task instance that produced the conclusion and the original data on which it is based. This achieves the interpretability, verifiability and traceability of the conclusions, and solves the problems of difficulty in attributing conclusions and the inability of users to verify the reliability of conclusions in traditional multi-agent systems.

[0186] See Figure 3 , Figure 3 A flowchart of another data processing method provided in one embodiment of this specification is shown.

[0187] Step 302: The task to be processed is processed using a multi-agent system, wherein the multi-agent system includes a scheduling agent and a communication coordination agent, and the sub-agents in the multi-agent system send task execution information to the scheduling agent.

[0188] Step 304: The scheduling agent receives the task execution information and forwards it to the communication coordination agent. If the communication coordination agent determines that the task execution information meets the broadcasting conditions, it converts the task execution information into natural language information and outputs it to the corresponding user client.

[0189] In the embodiments of this specification, a sub-agent can be understood as an agent that performs reasoning for a specific task. For example, a sub-agent can be a task execution agent that processes sub-tasks in the above implementation.

[0190] Specifically, the independent communication coordination agent is designated as the sole outlet for outputting information to the user, while the calling agent is used merely as a scheduler and cannot speak directly to the user. In practice, the broadcasting tool (which is used to send task execution information to the user) is registered with the communication coordination agent and can only allow messages from the communication coordination agent to pass through, thus ensuring that natural language information is sent to the user only through the communication coordination agent.

[0191] For specific implementation methods, please refer to the above embodiments, which are not limited here.

[0192] The data processing method provided in the embodiments of this specification avoids the problem of protocol mark leakage (users seeing incomprehensible marks such as [done], task_id, etc.) in the process reporting stage due to the mixing of technical language and natural language in the same context window by the scheduling agent, or the scheduling agent relaxing the precision of instructions in order to output natural language information. It can eliminate language pollution. Moreover, it reduces the burden on the scheduling agent by deleting the rules related to "speaking discipline" (speaking timing, wording norms, counterexamples, instructions for using broadcasting tools, etc.) in the instruction file of the scheduling agent and retaining only the scheduling rules. This makes the scheduling agent's scheduling behavior more stable and reliable when it only implements the scheduling function, and no longer causes abnormal behavior due to "inconsistencies between scheduling discipline and speaking discipline".

[0193] See Figure 4 , Figure 4 A flowchart illustrating the processing procedure of a data processing method according to an embodiment of this specification is shown, specifically including the following steps.

[0194] The data processing method provided in the embodiments of this specification introduces a complete chain of "safety valve access - task splitting - capability matching - dynamic generation - cross-questioning - acceptance and sedimentation - conclusion backtracking", and superimposes dynamic teaming and capability sedimentation on the existing predefined multi-agent system architecture to realize a closed loop from user task description to automatic multi-agent team formation, execution, quality assurance and capability accumulation.

[0195] In the embodiments of this specification, a basic intelligent agent and task instances are distinguished. The basic intelligent agent is a fixed execution shell used to provide process resources, model configuration, communication framework, and workspace. A task instance is a one-time responsibility perspective built upon the basic intelligent agent, with the focus direction distinguished by the task responsibility description. A single basic intelligent agent can simultaneously support multiple task instances executing in parallel, and the conclusions of each task instance can be cross-validated due to differences in perspective.

[0196] Step 402: Safety valve access determination.

[0197] Before a user task (i.e. a task to be processed) enters the multi-agent system for processing, it first undergoes an admission judgment by a safety valve agent (i.e., an admission judgment agent in the above embodiment) to determine whether the task to be processed is suitable for processing using the multi-agent collaborative mode, thereby preventing non-collaborative tasks from mistakenly entering the multi-agent processing link.

[0198] The safety valve agent is a pre-configured dedicated agent that uses an invisible session mechanism to perform judgments. This session is invisible in the system session list, does not trigger subsequent processing, is not written to the persistent session history, and does not generate diagnostic logs. It is automatically destroyed after the judgment is completed, achieving zero pollution of normal session history.

[0199] The safety valve agent's input is the user task, and the output is a binary result indicating whether to allow or block, along with a reason. The judgment criteria are: block if the information is severely incomplete or the task is clearly not a team collaboration task; otherwise, allow. A failure-open strategy is used when the judgment fails, allowing the task to enter the multi-agent collaboration process to avoid interrupting normal tasks due to safety valve malfunction.

[0200] When the safety valve agent outputs a blocking function, the user task is handled by the main agent in single-agent mode; when the output is allowed, the task proceeds to the subsequent task splitting stage.

[0201] Step 404: Task breakdown and capability requirement information generation.

[0202] When a user task passes through a safety valve agent, the task splitter decomposes the user task. The task splitter can be implemented using a large language model with a large number of parameters and strong reasoning capabilities, because task splitting is a complex reasoning process, requiring a regressive large language model to ensure the quality of the splitting.

[0203] The task splitter takes user tasks and related context as input and outputs a structured list of tasks and a capability requirement description. The capability requirement description uses business language to describe the capabilities required for each subtask. The task splitter also annotates sequential and parallel dependencies between subtasks, as well as dynamically expandable fan-out points.

[0204] The task splitter is designed according to the single responsibility principle, which means it only describes "what capabilities are needed". Therefore, the task splitter does not need to know the capabilities of the agent pool (i.e., the set of agents). This separation of responsibilities avoids the task splitter's bloated prompts and inaccurate judgments.

[0205] It should be noted that when a task is matched with a predefined team (i.e., the predefined intelligent agent system in the above embodiment), the user task can be processed directly through the predefined team. However, if no matching predefined team is found, this step and the subsequent dynamic team formation process will be triggered.

[0206] Step 406: Capability matching and dynamic generation of agents.

[0207] The capability matcher receives the capability requirement descriptions of each subtask generated by the task splitter and matches them against the existing pool of agents. For each subtask, if a suitable agent is matched, it is marked as reused; otherwise, it is marked as new.

[0208] The agent pool is a list of all registered agents in the system configuration, including pre-configured agents (i.e., pre-configured agents) and agents that have been solidified and converted (generated by solidifying agents in a temporary state). Each agent has an independent workspace, model configuration, tool permissions, and personality definition.

[0209] For newly created target subtasks, the capability matcher uses a large language model for reasoning. Based on the capability features described in the capability requirement description, it selects a suitable initial agent (i.e., basic agent) for the target subtask. It then uses template injection to inject personality configuration information and task instruction configuration information into the general initial agent to generate a temporary task execution agent.

[0210] The difference between template injection and generation from scratch is that generating an agent from scratch requires writing system prompts out of thin air using a large language model, which can easily deviate from communication norms. In contrast, the basic agent of template injection has a built-in communication protocol, and personality configuration information and task instruction configuration information are used to describe business capabilities.

[0211] A multi-agent system corresponding to the target task can be constructed by a scheduling agent and various task execution agents. It should be noted that a communication and coordination agent can also be injected into the multi-agent system.

[0212] Step 408: Execute the task instance.

[0213] Based on the task responsibility description of each subtask and the corresponding task execution agent, each task instance is formed, and each task instance is executed on its respective bound base agent.

[0214] The centralized scheduler, as a single global scheduling component, manages message queues by team and session identifiers. Messages within the same team and session are processed serially in strict first-in-first-out order, while messages from different sessions can be processed in parallel. The centralized scheduler maintains a persistent session registry that records the session binding relationships and lifecycle states of each task instance, and supports automatic recovery of active sessions after a process restart.

[0215] The persistent inbox serves as an asynchronous communication channel between agents. The scheduling agent sends execution instructions to the inbox of the target task instance through the job dispatch tool, and the task instance sends the conclusion back to the scheduling agent's inbox through the reporting tool. The inbox messages are persistently stored on the disk, including the sender, receiver, timestamp, and body, and are not lost after the process restarts.

[0216] During execution, independent subtasks are executed in parallel, while dependent subtasks wait sequentially for the output of upstream tasks. Multiple task instances of the same basic agent can execute different analysis aspects in parallel.

[0217] In this embodiment, the conclusion reporting of task instances adopts a dual-channel guarantee mechanism. The main route is for the agent to report the conclusion by calling the inbox to send the tool, and the fallback route is for the script to report the conclusion by checking the session record file. If the agent does not call the sending tool, the script will automatically forward the conclusion to ensure that the conclusion is not lost due to the omission of tool calls.

[0218] When a task instance produces a conclusion, it can attach a structured challenge message at the end of the conclusion and report it to the scheduling agent through the inbox.

[0219] Step 410: Cross-signal structured challenge.

[0220] When a task instance reports its conclusion to the scheduling agent, the scheduling agent scans the structured challenge information attached to the conclusion and executes the structured challenge process.

[0221] Structured questioning information is optional and can be attached to the task instance when it produces a conclusion. Specifically, it is in the form of a structured table and includes four fields: target role description (describes the target task instance’s ability domain and must be able to match the team’s role list), follow-up questions, data basis, and priority. Among them, the data basis is the quantitative basis that supports the follow-up questions.

[0222] Based on the target role description, the system matches the corresponding target task instance from the list of team roles included in the multi-agent team; that is, the role corresponding to the target task instance matches the target role description. Follow-up questions are sent to the target task instance via its inbox. Upon receiving these questions, the target task instance processes them as additional queries and then sends back its conclusion.

[0223] Compared to free group chat-style collaboration, structured questioning has the advantages of convergence and determinism, with the scheduling agent adjudicating instead of unlimited multi-turn dialogues. Each question and response is an independent inbox message, which is traceable and only targets the parts that need to be questioned, resulting in lower costs.

[0224] Step 412: Schedule intelligent experience reception.

[0225] After all task instances report their conclusions, the scheduling agent executes the acceptance process. The acceptance process includes the following steps: performing structured challenge information scanning and routing; checking the output of each task instance item by item according to the acceptance criteria preset by the task splitter, including whether the output contains all required fields, whether the data format meets the specifications, and whether the conclusion has traceable data support; finally, classifying the agent to which the task instance belongs based on the acceptance record, marking a pass as qualified, a pass after one revision as observation, and a failure to review, timeout, or reassignment as poor.

[0226] After acceptance, the scheduling agent delivers the conclusions of each task instance to the aggregation task instance, which integrates them into a final report (i.e., task processing results). When the aggregation task instance encounters conflicting conclusions, it reports to the scheduling agent without making its own decision.

[0227] Step 414: Precipitation turns positive.

[0228] For temporary intelligent agents that are rated as qualified in the acceptance process, a sedimentation operation is performed to convert them into permanent intelligent agents.

[0229] Specifically, the temporarily injected personality definition configuration information and task instruction configuration information are permanently stored in the workspace directory corresponding to the agent; then the agent is registered in a separate persistent sedimentation file, and the agent in the sedimentation file is automatically merged into the capability pool configuration when the system starts.

[0230] The sedimentation operation is implemented asynchronously. This means that the current task uses a temporarily injected agent, while the sedimented permanent agent enters the reusable capability pool after the next system restart. Subsequent similar tasks can then reuse this agent without regenerating it. This sedimentation mechanism achieves a closed loop of continuous growth in system capabilities.

[0231] It should be noted that the sedimentation files are stored using an independent file scheme rather than being directly written to the system configuration file, in order to avoid conflicts with the agent list rearrangement and completion logic during system startup.

[0232] Step 416: Conclusion backtracking.

[0233] In this embodiment, a three-layer backtracking capability enables each conclusion to be traced back from the final report to the agent that produced the conclusion and the original data on which it was based.

[0234] The first layer is the preservation of the architectural evidence chain: each task instance is an independent intelligent agent session with a complete dialogue record; the conclusion of the task instance is reported through a persistent inbox, and the inbox message contains the sender, receiver, timestamp and body, which is persistently stored and will not be lost; the output file can be persistently saved.

[0235] The second layer is interactive tracing: As a task dispatcher, the scheduling agent naturally knows which tasks were assigned to whom and who reported what. Therefore, when a user inquires about a certain conclusion, the scheduling agent can explicitly trace the source link of the conclusion, specifically from the conclusion to the corresponding task instance, then to the corresponding discovery, and then to the original data. In other words, the scheduling agent can actively trace the source based on its scheduling knowledge.

[0236] The third layer is dynamic re-investigation: After the scheduling agent traces back to the corresponding task instance, it can request the task instance to assist in further tracing. That is, backtracking is not a read-only "checking records", but can reactivate the corresponding agent for further analysis.

[0237] The data processing method provided in the embodiments of this specification realizes the automated construction of a multi-agent team with a complete closed loop from "task splitting, agent matching, template injection generation, and reuse". It also saves the qualified temporary agents into permanent agents and puts them into the capability pool, realizing the capability growth of "the more you use it, the stronger it becomes". When the task instance produces a conclusion with structured question information, the scheduled agent scans and routes to the target task instance for targeted questioning, replacing the free group chat-style collaboration. It has several advantages, including convergence and certainty, traceability (inbox messages), and low cost (targeted questioning instead of full discussion).

[0238] See Figure 5 , Figure 5 This specification illustrates a flowchart of a data processing method according to an embodiment of the present specification, which coordinates the interaction between an intelligent agent and a user through communication.

[0239] In the field of user interaction in multi-agent systems, the scheduling agent (main agent) in existing solutions usually undertakes two sets of responsibilities at the same time: "process control" and "speaking to users". The two sets of constraints interfere with each other in the same prompt word, causing the scheduling agent to remain silent when focusing on scheduling, and easily introduce protocol tags and analyst style when it needs to speak, which is incomprehensible to users.

[0240] When a scheduling agent needs to use both technical language for sub-agents and natural language for users in the same session context, maintaining two language styles simultaneously within the same context window presents structural difficulties and inevitably leads to mutual permeation, i.e., technical language leaks to users or user-faced language dilutes the precision of instructions.

[0241] Furthermore, the scheduling agent relies on "self-reporting" progress to users. In reality, through testing, it was found that the scheduling agent usually only sends a confirmation message at the beginning of the task and a final report at the end, remaining silent throughout the process, so users cannot understand the team's progress.

[0242] In the embodiments described in this specification, by introducing an independent communication coordination agent as the sole information outlet for the user, and using the scheduling agent as a simple scheduler, user interaction and internal scheduling are fundamentally decoupled. The following, in conjunction with... Figure 5 The data processing method provided in this embodiment will be described in detail below: Specifically, a communication and coordination agent is automatically injected when a multi-agent team is created. The injection occurs in a mandatory function of the team creation path; if the multi-agent team already contains a communication and coordination agent, it is not injected again.

[0243] As an independent agent, the communication coordination agent possesses independent personality configuration, skill modules, and conversation context. Regarding personality configuration, the communication coordination agent is defined as a broadcaster role, with its role clearly defined as "extracting but not analyzing," meaning it only extracts and paraphrases existing conclusions, without conducting independent business analysis, generating business documents, or participating in scheduling decisions.

[0244] In terms of skill configuration, a dedicated skill module can be loaded into the workspace of the communication coordination agent. This skill module defines the writing specifications for broadcasts, including fidelity rules, richness of expression, product path references, and question closings. By adding, deleting, or replacing this skill module and adjusting the personality definition, the output style can be plugged in and changed without modifying any configuration of the scheduling agent or task instance.

[0245] In terms of session configuration, the communication coordination agent has an independent session that belongs to the user interaction domain. This session includes natural language notifications, user-oriented explanations and reports, but does not contain any internal protocol tags or structured instructions. It is physically isolated from the session context of the scheduling agent, which only contains AI collaboration domain text, thus eliminating the mutual contamination between protocol / technical language and natural language at the context level.

[0246] This communication coordination agent is completely invisible to users. It skips all user-visible interactions in the system, so users are unaware of its existence. However, it is fully retained in the debugging interface for troubleshooting purposes.

[0247] During the execution of a task instance in a multi-agent system, when the scheduler delivers inbox messages (including task splitter plans, task instance conclusion reports, help signals, external fault reports, etc.) to the scheduling agent, the system automatically copies a complete set of messages to the inbox of the communication coordination agent.

[0248] The specific logic of event copying is as follows: the scheduling agent receives the original as usual, the scheduling process remains unchanged, and the communication coordination agent receives a complete copy, possessing an information field of equal importance to the scheduling agent; messages from the communication coordination agent itself are no longer copied to avoid message loops; the scheduler delivers messages according to the session queue, and multiple events are queued to enter the inbox of the communication coordination agent. Continuous events can be naturally accumulated in the session context of the communication coordination agent to achieve merged broadcasting.

[0249] It should be noted that during the copy operation, the communication coordinating agent for the team is searched for by role identifier in the session registry. If it cannot be found, it is silently skipped to ensure compatibility with older teams. The copy operation is triggered when the output of the scheduling agent session is completed, and the complete output text is delivered to the communication coordinating agent's inbox to ensure that the communication coordinating agent does not miss the output because it does not know what the scheduling agent replied.

[0250] When a message copied by an event arrives and is triggered, the communication coordination agent determines whether the information meets the broadcasting conditions. If the copied information belongs to a predefined broadcasting type such as opening confirmation, progress report, final delivery report, or requires user decision and has actual information value, it is determined to be output information. If there is no substantial progress content or no information increment, a non-broadcasting flag is returned and no output is performed this time.

[0251] For example, in a scenario where the scheduling agent actively dispatches tasks, after collecting and accepting all conclusions, the scheduling agent submits a "delivery broadcast order" to the communication coordination agent through the task dispatch tool. This order includes acceptance notes and a product path list. Upon receiving the order, the communication coordination agent reads the product file, generates a final delivery report, and broadcasts it. In a scenario where user decision-making is required, the scheduling agent submits a "decision request order" to the communication coordination agent. The communication coordination agent translates this into a user-readable decision request and broadcasts it to the user.

[0252] After determining the information to be output, the communication coordination agent converts the internal task execution information into natural language information with the target output style according to the target output style corresponding to its loaded exclusive skill module and personality definition. Then, the natural language information is verified to check whether it contains internal technical markers (such as status markers, task identifiers, session identifiers, etc.). If detected, it is replaced with a user-readable natural language expression to ensure that no internal jargon remains in the information output to the user.

[0253] After successful verification, the communication coordination agent invokes the user broadcasting tool to output natural language information to the user's corresponding client. The user broadcasting tool is the only legitimate way to send messages to users. The registration of this tool adopts an identity gating mechanism, that is, it is only registered on sessions where the communication coordination agent's role identifier is found in the session registry. Sessions with other roles are not registered with this tool, ensuring that the scheduling agent and task instances cannot speak directly to users.

[0254] Users interact with the multi-agent system through a chat window in the scheduling agent's conversation. This chat window is the only visible interface for the user within the entire system. Users input task requests in this chat window, and the system also uses it to return processing results and progress updates. The scheduling agent's conversation chat window is controlled by a whitelist at the display level; only genuine user messages, system messages, and communication coordination agent broadcast messages are allowed to be displayed to the user, while all other messages are filtered out.

[0255] The whitelist display is determined using a runtime fact-driven approach based on the presence of the communication coordination agent: a record of the communication coordination agent is checked in the session registry; if it exists, whitelist mode is enabled; otherwise, blacklist mode is reverted to ensure historical compatibility. In case of anomalies, blacklist mode is used as the default security measure. Through this dual-layer control of access control and whitelist display, users only perceive the natural language output of the communication coordination agent during tasks within the multi-agent system; the system's internal collaboration information is never directly exposed to the user.

[0256] In this embodiment, a multi-layered reliability guarantee mechanism is used to prevent the communication coordination agent from missing important information or making translation errors. For example, the system retains a fallback mechanism for missing final reports. That is, if the delivery broadcast from the communication coordination agent is not received within a preset time, a reminder is triggered to remind the scheduling agent to immediately dispatch a delivery broadcast order to the communication coordination agent. In addition, the communication coordination agent retains its historical broadcast records in the session context, thereby determining "the same progress should not be broadcast repeatedly" based on the context.

[0257] The data processing method provided in the embodiments of this specification achieves a single sound outlet by coordinating the intelligent agent with independent communication and a two-layer access control mechanism. This completely eliminates the confusion and information black box problem caused by multiple intelligent agents directly replying to users in traditional solutions. The amount of information perceived by the user has increased from an average of 2 messages to 6-7 progress reports with substantial content.

[0258] By isolating the AI ​​collaboration domain and the user interaction domain into independent sessions for different intelligent agents, the possibility of protocol tags leaking to the user interface is physically eliminated. Users only see pure natural language broadcasts throughout, with internal technical jargon completely blocked. At the same time, the scheduling agent no longer assumes any broadcasting responsibilities, and its behavior is more stable and reliable after all speech discipline rules are removed from the instruction file.

[0259] As an independent component, the communication coordination agent's output style can be hot-swapped by adding or deleting skill modules or adjusting personality definitions. It is completely orthogonal to the system's core scheduling logic, thus completely resolving the fundamental contradiction in traditional solutions where "changing styles harms scheduling, and changing scheduling harms styles." Furthermore, the communication coordination agent is completely invisible to the user and is compatible with historical conversations.

[0260] The complete event copying mechanism enables the communication coordination agent to obtain an information view equivalent to that of the scheduling agent, and to independently judge the timing and content of broadcasting without relying on the model's self-awareness; the deterministic event-driven arousal further transforms the timing judgment from unreliable model self-awareness into a reliable system-level guarantee.

[0261] By employing mechanisms such as omission prevention (event copy completion + final report fallback), error translation prevention (terminology verification + skill authenticity rules + tool whitelist + scheduling agent not commanding broadcast content), anti-spam (activity assignment copy can be switched on and off), and anti-duplication (session memory suppression), the communication coordination agent is ensured to avoid mis-broadcasting, omissions, leakage of internal protocol language, and spamming. This makes the communication coordination agent a reliable system-level component, improving the effectiveness of multi-agent collaboration.

[0262] One embodiment of this specification also provides a data processing system, including a scheduling agent, sub-agents, and a communication coordination agent, wherein... The scheduling agent and the sub-agent are used to process the received tasks to be processed. The sub-agent sends task execution information to the scheduling agent, and the task execution information is copied to the communication coordination agent. The communication coordination agent is used to convert the task execution information into natural language information and output it to the corresponding client of the user when it is determined that the task execution information meets the broadcasting conditions.

[0263] For specific implementation details, please refer to the above embodiments, which will not be repeated here.

[0264] Corresponding to the above method embodiments, this specification also provides data processing apparatus embodiments. Figure 6 A schematic diagram of the structure of a data processing apparatus according to one embodiment of this specification is shown. Figure 6 As shown, the device includes: The disassembly module 602 is configured to disassemble the target task and generate at least one subtask and the capability requirement information corresponding to each subtask. The determination module 604 is configured to determine the corresponding task execution agent for each subtask from the set of agents based on the capability requirement information corresponding to each subtask. The creation module 606 is configured to create a task execution agent corresponding to the target subtask based on the initial agent and the capability requirement information of the target subtask when there is no task execution agent corresponding to the target subtask in the agent set. The module 608 is configured to execute at least one subtask using a scheduling agent and a task execution agent corresponding to each subtask, and obtain the task processing result.

[0265] Optionally, the obtaining module 608 is further configured to: The scheduling agent and the task execution agent corresponding to each subtask are used to execute the at least one subtask. During the execution, the task execution information received by the scheduling agent is copied to the communication coordination agent. Once the communication coordination agent determines that the task execution information meets the broadcasting conditions, it converts the task execution information into natural language information and outputs it to the corresponding user client.

[0266] The device further includes: The receiving module is configured to receive tasks to be processed, determine the task type of the task to be processed using an admission judgment agent, and determine the task to be processed as a target task if the task type of the task to be processed is a target type, wherein the target type is the type of task that needs to be processed by multiple agents.

[0267] Optionally, the determining module 604 is further configured to: Determine the agent configuration information of each agent in the agent set, wherein the agent configuration information includes personality configuration information and task instruction configuration information; Semantically match the capability requirements of each subtask with the agent configuration information of each agent, and determine the matching agent for each subtask based on the semantic matching results.

[0268] Optionally, the determining module 604 is further configured to: If the semantic matching result between the capability requirement information corresponding to the subtask and the agent configuration information of any agent in the agent set meets a preset threshold, the corresponding agent is determined as the task execution agent corresponding to the subtask. Otherwise, mark the subtask as the target subtask that requires the creation of a new agent.

[0269] Optionally, the determining module 604 is further configured to: Based on the capability requirement information of the target sub-task, an initial intelligent agent is determined from the set of general intelligent agents, wherein each general intelligent agent in the set of general intelligent agents has different processing capabilities in different processing dimensions; Based on the role information and task instruction information in the capability requirement information corresponding to the target sub-task, determine the personality configuration information and task instruction configuration information, and determine the role toolset corresponding to the role information; The personality configuration information and task instruction configuration information are injected into the initial agent, and the role toolset is configured for the initial agent to obtain the task execution agent corresponding to the target sub-task.

[0270] Optionally, the obtaining module 608 is further configured to: Create corresponding task instances for each subtask, and bind each task instance to the corresponding task execution agent; Based on the dependencies between subtasks, the scheduling agent schedules each task instance.

[0271] Optionally, the obtaining module 608 is further configured to: Based on the dependencies between subtasks, determine the first type of subtasks to be executed in parallel and the second type of subtasks to be executed sequentially; For the first type of subtask, the scheduling agent schedules the corresponding task instances in parallel for execution. For the second type of subtask, the scheduling agent schedules the corresponding task instances in sequence according to the dependency order for execution. The scheduling agent sends execution instructions to the task instance through the inbox corresponding to the task instance, and the task instance returns task execution information to the scheduling agent through the inbox corresponding to the scheduling agent.

[0272] The device further includes: The interrogation module is configured to, when any task instance sends structured interrogation information to the inbox corresponding to the scheduling agent, use the scheduling agent to determine the target task instance to be interrogated, wherein the structured interrogation information includes a target role identifier and follow-up questions, and the target task instance is the task instance corresponding to the target role identifier. The scheduling agent sends follow-up questions to the target task instance through the inbox of the target task instance, and obtains the task execution information returned by the target task instance based on the follow-up questions.

[0273] The device further includes: The acceptance module is configured to accept the execution results of each task instance according to preset acceptance criteria, and obtain the acceptance results of each task instance; if the acceptance result is unsuccessful, a modification instruction is sent to the corresponding task instance, the modified execution result is received, and acceptance is performed again; after the execution results of each task instance pass the acceptance, the execution results are integrated to obtain the task processing result.

[0274] The device further includes: The solidification module is configured to determine the quality level of the task execution agent corresponding to each target subtask based on the acceptance record of the execution result corresponding to each task instance; if the quality level is qualified, the task execution agent corresponding to the target subtask is solidified as the target execution agent and stored in the agent set.

[0275] The device further includes: The tracing module is configured to save the execution session records, inbox messages and execution results of each task instance to form an evidence chain; in response to a tracing request for the target conclusion in the task processing result, it determines the task instance that produced the target conclusion and the conclusion basis corresponding to the target conclusion based on the evidence chain.

[0276] The device further includes: The judgment module is configured to determine whether there is a predefined intelligent agent system matching the target task based on the task description information of the target task; if there is, the target task is processed using the predefined intelligent agent system; if not, the step of decomposing the target task is executed.

[0277] The device further includes: The output module is configured to convert the task execution information into natural language information with the target output style according to the target output style corresponding to the communication coordination agent; verify whether the natural language information contains internal technical tags, and if so, replace the internal technical tags with natural language information with the target output style; and use the communication coordination agent to call a broadcasting tool to output the verified natural language information to the user's corresponding client.

[0278] Optionally, the output module is further configured as follows: When the natural language information originates from the communication coordination agent, the natural language information is output to the user's corresponding client.

[0279] Optionally, the output module is further configured as follows: If the communication coordination agent determines that the task execution information belongs to a predefined broadcast type, it determines that the task execution information meets the broadcast conditions.

[0280] The above is an illustrative scheme of a data processing apparatus according to this embodiment. It should be noted that the technical solution of this data processing apparatus and the technical solution of the data processing method described above belong to the same concept. For details not described in detail in the technical solution of the data processing apparatus, please refer to the description of the technical solution of the data processing method described above.

[0281] Corresponding to the above method embodiments, this specification also provides another data processing apparatus embodiment, which includes: The processing module is configured to process the task to be processed using a multi-agent system, which includes a scheduling agent and a communication coordination agent, wherein each sub-agent in the multi-agent system sends task execution information to the scheduling agent. The output module is configured to copy the task execution information received by the scheduling agent to the communication coordination agent, and, if the communication coordination agent determines that the task execution information meets the broadcasting conditions, convert the task execution information into natural language information and output it to the corresponding user client.

[0282] The above is an illustrative scheme of a data processing apparatus according to this embodiment. It should be noted that the technical solution of this data processing apparatus and the technical solution of the data processing method described above belong to the same concept. For details not described in detail in the technical solution of the data processing apparatus, please refer to the description of the technical solution of the data processing method described above.

[0283] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0284] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0285] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0286] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 700 can also be a mobile or stationary server.

[0287] The processor 720 is used to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the above-described data processing method.

[0288] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the computing device embodiments are basically similar to the data processing method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the data processing method embodiments.

[0289] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described data processing method.

[0290] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the computer-readable storage medium embodiments are basically similar to the data processing method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the data processing method embodiments.

[0291] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described data processing method.

[0292] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the data processing method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the data processing method described above.

[0293] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0294] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0295] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0296] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0297] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A data processing method, comprising: The target task is broken down into tasks, generating at least one subtask and the capability requirements information for each subtask. Based on the capability requirements of each subtask, a corresponding task execution agent is determined from the set of agents for each subtask. If there is no task execution agent corresponding to the target subtask in the set of agents, a task execution agent corresponding to the target subtask is created based on the initial agent and the capability requirement information of the target subtask. The scheduling agent and the task execution agent corresponding to each subtask are used to execute at least one subtask to obtain the task processing result.

2. The method as described in claim 1, wherein the at least one subtask is executed using a scheduling agent and a task execution agent corresponding to each subtask to obtain a task processing result, further comprising: The scheduling agent and the task execution agent corresponding to each subtask are used to execute the at least one subtask. During the execution, the task execution information received by the scheduling agent is copied to the communication coordination agent. Once the communication coordination agent determines that the task execution information meets the broadcasting conditions, it converts the task execution information into natural language information and outputs it to the corresponding user client.

3. The method as described in claim 1, before decomposing the target task and generating at least one sub-task and corresponding capability requirement information for each sub-task, further includes: Receive tasks to be processed and use an admission judgment agent to determine the task type of the tasks to be processed; If the task type of the task to be processed is a target type, the task to be processed is determined as a target task, wherein the target type is the type of task that needs to be processed by multiple agents.

4. The method as described in claim 1, wherein, based on the capability requirement information corresponding to each sub-task, a corresponding task execution agent is determined from the set of agents for each sub-task, comprising: Determine the agent configuration information of each agent in the agent set, wherein, The intelligent agent configuration information includes personality configuration information and task instruction configuration information; Semantically match the capability requirements of each subtask with the agent configuration information of each agent, and determine the matching agent for each subtask based on the semantic matching results.

5. The method as described in claim 4, wherein a matching task execution agent is determined for each subtask based on the semantic matching result, comprising: If the semantic matching result between the capability requirement information corresponding to the subtask and the agent configuration information of any agent in the agent set meets a preset threshold, the corresponding agent is determined as the task execution agent corresponding to the subtask. Otherwise, mark the subtask as the target subtask that requires the creation of a new agent.

6. The method as described in claim 1, wherein creating a task execution agent corresponding to the target sub-task based on the initial agent and the capability requirement information of the target sub-task, comprises: Based on the capability requirement information of the target sub-task, an initial intelligent agent is determined from the set of general intelligent agents, wherein each general intelligent agent in the set of general intelligent agents has different processing capabilities in different processing dimensions; Based on the role information and task instruction information in the capability requirement information corresponding to the target sub-task, determine the personality configuration information and task instruction configuration information, and determine the role toolset corresponding to the role information; The personality configuration information and task instruction configuration information are injected into the initial agent, and the role toolset is configured for the initial agent to obtain the task execution agent corresponding to the target sub-task.

7. The method of claim 1, wherein the at least one subtask is executed using a scheduling agent and a task execution agent corresponding to each subtask, comprising: Create corresponding task instances for each subtask, and bind each task instance to the corresponding task execution agent; Based on the dependencies between subtasks, the scheduling agent schedules each task instance.

8. The method as described in claim 7, wherein, based on the dependencies between subtasks, the scheduling agent schedules each task instance to execute the corresponding subtask, comprising: Based on the dependencies between subtasks, determine the first type of subtasks to be executed in parallel and the second type of subtasks to be executed sequentially; For the first type of subtask, the scheduling agent schedules the corresponding task instances in parallel for execution. For the second type of subtask, the scheduling agent schedules the corresponding task instances in sequence according to the dependency order for execution. The scheduling agent sends execution instructions to the task instance through the inbox corresponding to the task instance, and the task instance returns task execution information to the scheduling agent through the inbox corresponding to the scheduling agent.

9. The method of claim 8, further comprising: If any task instance sends a task execution message to the inbox corresponding to the scheduling agent containing structured challenge information, the scheduling agent is used to determine the target task instance being challenged. The structured challenge information includes a target role identifier and follow-up questions. The target task instance is the task instance corresponding to the target role identifier. The scheduling agent sends follow-up questions to the target task instance through the inbox of the target task instance, and obtains the task execution information returned by the target task instance based on the follow-up questions.

10. The method of claim 1, further comprising obtaining the task processing result: The execution results of each task instance are accepted according to the preset acceptance criteria to obtain the acceptance results of each task instance. If the acceptance result is unsuccessful, a modification instruction is sent to the corresponding task instance, the modified execution result is received, and the acceptance is performed again. After the execution results of each task instance have passed acceptance, the execution results are integrated to obtain the task processing result.

11. The method of claim 10, further comprising, after obtaining the task processing result: Based on the acceptance records of the execution results corresponding to each task instance, determine the quality level of the task execution agent corresponding to each target subtask; If the quality level is qualified, the task execution agent corresponding to the target subtask is solidified as the target execution agent and stored in the agent set.

12. The method of claim 1, further comprising, after obtaining the task processing result: Save the execution session records, inbox messages, and execution results of each task instance to form a chain of evidence; In response to a request to trace the source of the target conclusion in the task processing result, the task instance that produced the target conclusion and the conclusion basis corresponding to the target conclusion are determined according to the chain of evidence.

13. The method of claim 1, further comprising, before decomposing the target task: Based on the task description information of the target task, determine whether there is a predefined intelligent agent system that matches the target task; If it exists, the target task is processed using the predefined intelligent agent system; If it does not exist, perform the steps described above for task decomposition of the target task.

14. The method of claim 2, wherein converting the task execution information into natural language information and outputting it to the corresponding user client, includes: Based on the target output style corresponding to the communication coordination agent, the task execution information is converted into natural language information with the target output style; Verify whether the natural language information contains internal technical tags; if it does, replace the internal technical tags with natural language information in the target output style. The communication coordination agent invokes a broadcasting tool to output verified natural language information to the user's corresponding client.

15. The method of claim 14, further comprising outputting verified natural language information to the user's corresponding client, and including: When the natural language information originates from the communication coordination agent, the natural language information is output to the user's corresponding client.

16. The method of claim 14, wherein determining that the task execution information meets the broadcasting conditions includes: If the communication coordination agent determines that the task execution information belongs to a predefined broadcast type, it determines that the task execution information meets the broadcast conditions.

17. A data processing method, comprising: A multi-agent system is used to process the task to be processed. The multi-agent system includes a scheduling agent and a communication coordination agent. The sub-agents in the multi-agent system send task execution information to the scheduling agent. The scheduling agent receives the task execution information and forwards it to the communication coordination agent. If the communication coordination agent determines that the task execution information meets the broadcasting conditions, it converts the task execution information into natural language information and outputs it to the corresponding user client.

18. A multi-agent system, comprising a scheduling agent, sub-agents, and a communication and coordination agent, wherein, The scheduling agent and the sub-agent are used to process the received tasks to be processed. The sub-agent sends task execution information to the scheduling agent, and the task execution information is copied to the communication coordination agent. The communication coordination agent is used to convert the task execution information into natural language information and output it to the corresponding client of the user when it is determined that the task execution information meets the broadcasting conditions.

19. A data processing apparatus, comprising: The decomposition module is configured to decompose the target task and generate at least one subtask and the capability requirements information corresponding to each subtask. The determination module is configured to determine the corresponding task execution agent for each subtask from the set of agents based on the capability requirement information corresponding to each subtask; The creation module is configured to create a task execution agent corresponding to the target subtask based on the initial agent and the capability requirement information of the target subtask if no task execution agent corresponding to the target subtask exists in the set of agents. The acquisition module is configured to use a scheduling agent and a task execution agent corresponding to each subtask to execute at least one subtask and obtain the task processing result.

20. A data processing apparatus, comprising: The processing module is configured to process the task to be processed using a multi-agent system, which includes a scheduling agent and a communication coordination agent, wherein each sub-agent in the multi-agent system sends task execution information to the scheduling agent. The output module is configured to copy the task execution information received by the scheduling agent to the communication coordination agent, and, if the communication coordination agent determines that the task execution information meets the broadcasting conditions, convert the task execution information into natural language information and output it to the corresponding user client.

21. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the data processing method according to any one of claims 1-17.

22. A computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the data processing method according to any one of claims 1-17.

23. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the data processing method according to any one of claims 1-17.

Images